This PDF file contains the front matter associated with SPIE Proceedings Volume 12463, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Image Science provides a framework for the task-based assessment of image quality. This framework has been used to support the evaluation of medical imaging system hardware, iterative reconstruction algorithms and other image processing methods, and display devices by academia, industry, and US FDA. Since the earliest instances of the SPIE Medical Imaging Symposium the conference has served as an essential venue for presentations and discussions related to the objective assessment of image quality, featuring first disclosures of new models for physiological backgrounds and pathologies, tools for simulating medical imaging systems, models for the human and Bayesian observer, and methods for computing task-based figures of merit. This paper highlights recent advances in the objective or task-based assessment of image quality through the use of computational models and methods, and points to new initiatives intended to develop resources to move this important field forward.

Photon-counting CT (PCCT) is an emerging imaging technology with potential improvements in quantification and rendition of micro-structures due to its smaller detector sizes. The aim of this study was to assess the performance of a new PCCT scanner (NAEOTOM Alpha, Siemens) in quantifying clinically relevant bone imaging biomarkers for characterization of common bone diseases. We evaluated the ability of PCCT in quantifying microarchitecture in bones compared to conventional energy-integrating CT. The quantifications were done through virtual imaging trials, using a 50 percentile BMI male virtual patient, with a detailed model of trabecular bone with varied bone densities in the lumbar spine. The virtual patient was imaged using a validated CT simulator (DukeSim) at CTDIvol of 20 and 40 mGy for three scan modes: ultra-high-resolution PCCT (UHR-PCCT), high-resolution PCCT (HR-PCCT), and a conventional energy-integrating CT (EICT) (FORCE, Siemens). Further, each scan mode was reconstructed with varying parameters to evaluate their effect on quantification. Bone mineral density (BMD), trabecular volume to total bone volume (BV/TV), and radiomics texture features were calculated in each vertebra. The most accurate BMD measurements relative to the ground truth were UHR-PCCT images (error: 3.3% ± 1.5%), compared to HR-PCCT (error: 5.3% ± 2.0%) and EICT (error: 7.1% ± 2.0%). UHR-PCCT images outperformed EICT and HR-PCCT. In BV/TV quantifications, UHR-PCCT (errors of 29.7% ± 11.8%) outperformed HR-PCCT (error: 80.6% ± 31.4%) and EICT (error: 67.3% ± 64.3). UHR-PCCT and HR-PCCT texture features were sensitive to anatomical changes using the sharpest kernel. Conversely, the texture radiomics showed no clear trend to reflect the progression of the disease in EICT. This study demonstrated the potential utility of PCCT technology in improved performance of bone quantifications leading to more accurate characterization of bone diseases.

In recent years, deep learning-based methods for image denoising have drawn increasing attention and shown promising performance in denoising for low-dose CT images. Typically, deep learning utilizes paired data of low-dose images as the input and high-quality images as the target. In clinical imaging, it is difficult to obtain such paired data due to dose concerns. As an effective counterpart, the Noise2Noise framework can train a network based on noisy images and achieve equivalent performance to a conventional training strategy using high-quality images, namely Noise2Clean. In this work, we demonstrate the effectiveness of the Noise2Noise framework for denoising photon counting CT images. The noisy pairs can be generated from existing scans by splitting the raw counts into two scans with independent noise. Photon counting detectors count the number of incident photons, which follow a Poisson distribution. Binomial selection is applied to these measurements to generate noisy pairs at desired dose levels for Noise2Noise training. We show that Noise2Noise-based training approaches the performance of Noise2Clean training through a simulation study. Additionally, increased diversity in dose levels during training improved the performance of Noise2Noise in handling ultra low-dose images.

Current diagnostic techniques detect osteoarthritis too late for non-surgical treatment. This study examines the use of contrast-enhanced photon-counting spectral computed tomography (PC-SCT) for diagnosis of joint conditions. The multi-material decomposition capability and high-resolution imaging of PC-SCT with a novel bismuth-contrast agent allowed imaging of cartilage and bone structure, information on the volumetric distribution of proteoglycan and water composition in cartilage, and calcium deposits in bone for improved joint diagnosis. With PC-SCT, osteoarthritis can be monitored and treated by imaging the entire joint structure and components in one scan.

In this work, we studied how the scatter-to-primary ratio changes with the energy threshold in photon counting CT. Both Monte Carlo simulations and experimental measurements demonstrated a significantly reduced SPR in the high-energy bin projection data and reduced scatter artifacts in the reconstructed CT images. A new data acquisition scheme named Mohi-Polo was proposed by pushing the high energy threshold towards maximum keV, which generates a quasi-Monochromatic high energy bin and a Polychromatic low energy bin. Utilizing the nearly scatter-free high-energy bin data, we developed an innovative solution to correct the scatter artifacts in the low-energy bin.

Purpose: To investigate spectral distortion caused by bowtie (BT) scatter in photon-counting CT (PCCT) and its impact of the accuracy of material decomposition.

Methods: GPU-accelerated Monte Carlo (MC) simulations of a PCCT scanner with 1 mm detector pixels, a 1- dimensional anti-scatter grid (ASG) with 30:1 grid ratio, and five energy channels per pixel (25/35/50/65/80keV) were performed. Beam collimation was varied 80 – 160 mm (at isocenter). X-ray tube voltage was 120 kVp. An ~80 mm thick Aluminum BT was placed 48 mm from the x-ray focal spot. Spectral and spatial distributions of scatter and Scatter-to-Primary ratio (SPR) of water cylinders (150 mm – 300 mm diameter) were compared for two MC simulation settings: (a) scatter occurs both in the BT and in the object (ground-truth); and (b) BT acts only as a beam shaper but does not cause scatter. Water cylinders with Calcium wedge inserts (5 mm – 25 mm Ca thickness) were used to evaluate material decomposition errors due to object and BT scatter using least-squares projection-domain decomposition.

Results: Even with an 1D ASG, the low-energy channel SPRs for large objects (300 mm diameter) might be as high as 5 (at 80 mm collimation) – 10 (160 mm collimation). When BT scatter is ignored in scatter modeling, the SPRs are underestimated by 10% - 30%, with the relative error increasing with channel energy. Scatter introduces substantial biases in material decomposition: without any correction, the relative errors of Ca path length estimates range from 6%-10% for a 150 mm object and 80 mm collimation to 50%-100% for a 300 mm object at 160 mm collimation. Approximate correction using only object scatter reduces these biases to generally <5%, except for the largest phantom and collimation, where ignoring BT scatter leads to 6% - 12% underestimation of Ca thickness.

Conclusions: Algorithmic scatter corrections will likely be necessary for precise PCCT material decomposition for body-sized objects and wide collimations. While adequate results can be obtained using simplified models that ignore BT scatter, achieving <10% decomposition accuracy might require incorporating BT scatter.

We investigate the application of stationary Digital Tomosynthesis (DTS) using distributed x-ray source arrays for retrospectively-gated dynamic chest imaging under free-breathing conditions. The new capability might provide improved assessment of lung function impairment. A high-fidelity polychromatic x-ray system model was used to simulate a dynamic DTS configuration consisting of a linear array of 42 sources covering a 27.2° angular range in the superior-inferior direction. The array was placed 130 cm from a 42x42 cm flat-panel detector (FPD) with 0.56 mm pixels. The object center was placed at 116 cm from the source array. Dynamic acquisitions of a deformable digital chest phantom undergoing a realistic respiratory cycle (4 sec period) were simulated. The x-ray sources in the array were sequentially rastered at 8 fps; multiple passes through the array were performed to cover 3-19 breathing cycles. For reconstruction (0.5 mm voxels) of a given respiratory phase, a short sequence of projections temporally centered on the phase of interest was extracted from each cycle. The length of this sequence was varied from 1 (exact gating) to 5 frames. We investigated tradeoffs in motion blur (worse with longer gating), sampling artifacts (better with more breathing cycles), and noise, assuming a fixed total scan dose of 20 mAs regardless of the number of source array passes. A DTS acquisition over 15 respiratory cycles with exact gating yields breathing phase reconstructions at inspiration, expiration, and mid-cycle that recover >70% of the contrast in a static reconstruction and achieve adequate suppression of undersampling artifacts. For expiration and expiration, wide gating with 3 views/cycle can be used to reduce noise and improve sampling without introducing appreciable motion blur (>65% contrast recovery). For intermediate respiratory phases, wide gating causes motion blur that may require algorithmic compensation. Dynamic lung imaging with stationary DTS is feasible with approx. 1 min scan time and retrospective gating.

Grating-based X-ray phase-contrast imaging (XPCi) systems offer higher sensitivity compared to other XPCi methods; however, realizing a high-resolution, compact, dose-efficient imaging system has been a significant challenge from technological and practical points of view until now. X-ray gratings quality and characteristics directly determine the final imaging quality, where a proper grating fabrication process can potentially minimize image artifacts and increase the system visibility. To achieve a high-resolution, compact, dose-efficient XPCi system, high-resolution detectors and high-resolution X-ray absorption gratings are a must. Moreover, an efficient image processing method is required to retrieve multimodal XPCi information—transmission, refraction (phase-contrast), and dark-field—efficiently and simultaneously. In this work, we report on a compact XPCi system that enables multimodal information retrieval through single-shot imaging with two-directional sensitivity. We first present an elegant cost-effective fabrication method to make high-resolution micropillar-based X-ray absorption gratings. A prototype 2D grating is fabricated with micropillars with 4 μm in diameter, periodicity of 16.3 μm, and aspect ratio of more than 40. This grating is then employed along with a prototype hybrid a-Se/CMOS direct conversion high-resolution X-ray detector with a pixel pitch of 8 μm in a compact system with a polychromatic microfocus source to perform X-ray phase-contrast imaging of various samples. We successfully demonstrate a single-shot XPCi and retrieve multimodal XPCi data with transmission and dark-field metrics. The final system is a promising candidate for XPCi applications as it facilitates single-shot imaging, which reduces the exposure dose on samples and yields multimodal XPCi images efficiently, all in a compact bench-top setup. The delivered X-ray dose at the sample, resolution of the system, and compactness of the reported imaging setup are potentially beneficial for ex vivo, in vivo, and computed tomography (CT) imaging applications.

Talbot-Lau grating interferometry (GI) can conduct X-ray phase contrast imaging outside synchrotron facilities and simultaneously provide attenuation, differential phase contrast, and small angle scattering information about imaging samples, where periodically distributed X-ray line sources are required. In this study, we proposed a novel cold-cathode flat-panel X-ray source with micro-array anode target to generate such line sources without using grating G0. Its cathode was composed of densely arranged zinc oxide (ZnO) nanowires, which can generate electrons by field electron emission effect, whereas micro periodic distributed Al-Mo-Al strips were utilized as anode target. X-ray spatial distribution and spectrum of the source with different anode target period (p0) were studied via using EGSnrc. Structured X-ray illumination required for GI was obtained under Mo strips. Mean contrast of the X-ray spatial distribution under the Mo strips of flat panel sources with p0 of 3, 15, 30, 126 m were 33.36%, 81.98%, 91.55%, and 98.63%, respectively. The source can eliminate the use of G0 and thus related limitations of G0 on Talbot-Lau GI.

Our aim was to quantify the effects of pulse pileup on single- and multi-energy CT images from a commercial PCD-CT. A multi-energy CT phantom with 10 inserts was scanned at tube currents across the entire available range. Mean CT numbers were recorded and the CT number and noise ratios between low and high bin images were calculated. Results demonstrated that pulse pileup does not create noticeable artifacts or cause major CT number deviations with the variation in CT numbers being less than 10 HU for all images evaluated, indicating minimal impact on single- and multi-energy performance in current clinical PCD-CT.

Semiconductor-based photon counting detectors (PCDs) measure each incident photon energy through direct conversion process. Comparing to the conventional scintillator-based energy-integrating detectors (EIDs), it can be made with smaller pixel sizes and hence improve the system spatial resolution. The actual performance of the PCDs in CT system is compromised from ideal mainly due to two factors: charge sharing effect and pulse pileup. In particular, the charge sharing effect introduces signal crosstalk between neighboring pixels and degrade the detection spatial resolution. In this study, we derive a rigorous charge sharing detection formalism based on a comprehensive detector response model. The simulated results are compared with a pixel-to-pixel covariance measurement from a CZT-based prototype photon counting system. The results suggests that while a large fraction of detected events is affected by charge sharing, the actual measured crosstalk between neighboring pixels is greatly suppressed by a proper detection threshold in the counting mode. To understand the practical impact in image quality and optimize system design, a simplified crosstalk model based on the estimated charge sharing event rate is integrated into a system level simulation framework with realistic system specifications. The simulated image MTFs are measured for systems with four different pixel sizes. Results indicate that with realistic charge sharing effect, in the pixel size range that we tested, the image MTF steadily increases as the detector pixel size decreases.

Si has multiple attractive semiconductive properties to be used as a radiation sensor material in photon counting detectors. Additionally, the abundance and mature manufacturing technologies of Si have made high-purity Si readily available at low costs. Among physical interactions between x-rays and Si, the majority are expected to be Compton and Rayleigh scatterings due to the low atomic number of Si. In this work, the percentage and destinies of scattered x-ray photons were quantified both experimentally and numerically. The results have guided us to optimize the Si PCD configuration for more efficient utilization of the side-scattered photons.

Photon counting detectors (PCDs) with energy discrimination capabilities provide numerous benefits over conventional energy integrating detectors (EIDs). PCDs enable higher spatial resolution, improved contrast-to-noise ratio, reduced beam hardening artifact, and quantitative material-selective imaging. However, the increased spatial resolution and multiple energy bins (e.g., 8 bins for a prototype photon counting CT using silicon-based PCDs) greatly increase the amount of data generated. As a consequence, projection data transmission from the detector to the processing computer becomes more challenging due to the limited bandwidth of the slip ring. In this work, we compare the performance of four projection-domain energy bin compression strategies: conventional bins, summed bins, binary weights, and continuous weights, using the raw projection data from a prototype photon counting CT using silicon-based PCDs.

We reduce the 8 energy bins of the projection data from the prototype silicon-based PCDs to 2 or 3 virtual measurements using the phantom independent and globally applicable bin compression strategies, which were each optimized by minimizing the Cramér–Rao lower bound (CRLB) of the virtual measurements using only material decomposition calibration data over a predefined material space, and no other a priori knowledge. We evaluate the performance of the above 4 bin compression strategies in reconstructed images of a Catphan700 phantom. The results show that the 2 measurements generated with continuous weights can provide comparable material decomposition (MD) and virtual monoenergetic images (VMI) that exhibit low bias- and near zero variance-penalty, to that of the original binned counts, with a data reduction of 75%. On the other hand, neither conventional bins, summed bins, nor binary weights can provide comparable results versus continuous weights, even if we use 3 virtual measurements. We also conclude that to achieve low bias and variance in MD and VMI images with only two measurements, it is necessary to first measure as many energy bins as possible, then use continuous weights to compress the binned data.

One of the challenges for photon counting detector (PCD)-based computed tomography (CT) is spectral distortion due to charge sharing (CS). We recently proposed multi-energy inter-pixel coincidence counter (MEICC), which uses energy-dependent coincidence counters (CCs) between each PCD pixel and its neighboring pixels and records coincident events via charge sharing during data acquisition. Previous studies have shown that the performance of MEICC is as good as other technologies including digital count summing (DCS) scheme. In this study, we develop an algorithm that uses MEICC PCD output to create CS-corrected data. The performance of the method is assessed using a Monte Carlo (MC) simulation.

Photon-counting detectors (PCDs) have many advantages over the energy-integrating detectors. To fully exploit the advantages, however, it is important to develop a method that characterizes the PCD data by considering major spectral distortion factors: charge sharing (CS) and pulse pile-up (PP). Two major challenges were remaining model–data mismatch and pixel-to-pixel deviations from “good pixels.” In this work, we proposed a PCD model that uses global count-rate-dependent model parameters to address the global model–data mismatch problem and local parameters to handle local pixel-to-pixel deviations. The agreement of the proposed model and physical PCD data was outstanding.

Cadmium telluride (CdTe) is one of the materials used in photon-counting detectors for x-ray computed tomography. One challenge with this material is that it is susceptible to polarisation due to holes being trapped in impurities in the material. This can potentially lead to the buildup of bulk charge in the semiconductor, causing decreased charge collection efficiency and degraded energy resolution.

In this work, we develop a simulation model of CdTe detectors with polarisation and use it to study the effect of polarisation on the measured energy spectrum for different charge collection times. To this end, we use a theoretical model of charge buildup to find the critical charge in the detector’s bulk above which the detector can be considered completely polarised. We then simulate a 320-by-270-by-1600 μm CdTe detector used in CT clinical imaging, for varying degrees of polarisation (ratio between the actual charge and the critical charge) and charge collection time. Our results show that the measured spectrum gets heavily distorted for large degrees of polarisation or for short charge collection time. We also put these results in context by discussing how they relate to the critical fluence rate and the time of flight of the charge carriers. These results can lead to improved simulation models of CdTe detectors and a better understanding the factors affecting their imaging performance.

A 4D digital phantom model for contrast-enhanced breast imaging was developed in this work. The model is based on an anthropomorphic digital breast 3D phantom, which gives the user good control of the 4D phantom for developing and optimizing contrast enhancement imaging techniques. The fourth dimension of the phantom features different time-varying enhancement patterns for different materials including fibro-glandular tissue and lesion tissue. Physiological parameters that capture the key characteristics of masses, for example, wash-in and wash-out rates¹ and indications of metabolic level, are employed in the model to simulate certain fundamental features for categorizing mass types. The two-compartmental model,² a well-known model in the field of Pharmacokinetics, is used to depict the diffusion process of the contrast agent. Two methods are proposed to allow for the simulations of lesions with necrotic cores of varying shapes and sizes. With the contrast agent in the tissue, higher contrast is achieved between fibro-glandular tissue and mass tissue according to the mammograms simulated by Monte Carlo method.³

Coherence-based ultrasound imaging has demonstrated potential to improve breast mass diagnosis by distinguishing solid from fluid-filled masses. Harmonic imaging, which is known to reduce acoustic clutter, has the potential to offer additional improvements. However, the lack of a theoretical basis to describe these improvements precludes clinical recommendations based on physics and engineering principles. This work is the first to develop a theoretical model of coherence-based ultrasound imaging to describe both solid vs. fluid mass distinction and the effects of harmonic short-lag spatial coherence (SLSC) imaging. The scattering function and the transmit ultrasound beam of the van Cittert-Zernike theorem applied to ultrasound imaging were redefined to generate the theoretical model for solid vs. fluid mass distinction and for harmonic imaging, respectively. The derived theory was used to compare fundamental and harmonic SLSC images for hypoechoic solid, hypoechoic fluid, hyperechoic, and point targets. Theoretical simulations showed improved resolution, mitigated dark-region artifacts around hyperechoic targets, and increased spatial coherence of fluid masses in harmonic SLSC images when compared to fundamental SLSC images. Experimental data from tissue-mimicking phantoms and in vivo breast ultrasound images agreed with theoretical results. In particular, when compared to fundamental SLSC imaging, harmonic SLSC imaging improved resolution by 0.19 ± 0.25 mm, mitigated dark region artifacts by 0.55 ± 0.54 mm, and increased the spatial coherence of fluid-filled masses, resulting in a 6.50 ± 4.28 dB decrease in contrast. Results will enable future clinical recommendations supporting the use of fundamental or harmonic SLSC imaging for analyses of fluid or solid masses, respectively. These contributions establish a theoretical foundation to combine fundamental and harmonic coherence-based imaging with harmonic B-mode imaging to improve the accuracy of breast mass diagnoses.

Objective task-based assessment using physical phantoms,¹ and virtual clinical trials (VCTs)² are two useful approaches for evaluating new breast imaging technologies. Previous studies using these evaluation methods have suggested that lowering the radiation dose of mammograms would not lead to a significant change in the performance of the mass detection task.³ It has been speculated that decreases in the detectability of extended masses are mostly due to the level of anatomical structure commonly found in the breast, and less on the quantum noise of the system (and subsequently on the radiation dose level).⁴ However, these previous studies referred to above typically evaluate lesion detection performance using lobular mass models. The task of a breast radiologist in reading a mammogram is more complex than detection tasks usually modeled with physical phantom studies or VCTs.The radiologist typically searches throughout the breast for a suspicious lesion, and then determines a probability of malignancy for that lesion. This is also the task used for clinical reader studies designed to assess new technologies. Essentially this involves first a detection task which is then followed by a classification task. The classification of a malignant mass is sometimes dependent on the presence of high-frequency spiculations on the surface of the mass lesion. Visualization of mass spiculations are important because this suggests a higher chance to be malignant compared to a lobular mass. In this study, we use simulation studies to explore the use of a mass classification task to assess new breast technologies, instead of the often-used detection task. It is observed that unlike previous physical phantom and VCT experiments that assess detection of lobular masses, classification of spiculated masses does seem to be dependent on radiation dose level.

Cone-beam breast CT (bCT) provides volumetric images of the uncompressed breast but present higher noise than 2D mammography. Deep Learning (DL) denoising with supervised training has shown successful CBCT noise reduction but requires matched low-dose and high-dose images. Self-supervised training removes that requirement but often assume locally independent noise. This work studies the impact of bCT noise correlation on self-supervised denoising methods. The self-supervised training strategies included two blind spot methods – Noise2Self, enforcing local similarity with independent image noise; and Noise2Sim, enforcing image similarity in presence of correlated noise – and two Noisier2Noise approaches: i) noise injection in the image domain; and, ii) noise injection in projection domain with a model of noise correlation. Self-supervised training was performed on bCT images generated from 150 voxelized models with a high-fidelity forward projector, including models of the x-ray spectrum, polychromatic attenuation, and detector signal and noise propagation. Denoised images were assessed with respect to high-dose references and supervised denoising, using RMSE, SSIM, and noise power spectrum (NPS). Noise2Sim and Noisier2Noise with noise injection in the projection domain showed good performance in presence of correlated noise, achieving RMSE of 0.21 and 0.18 (SSIM of 0.9 and 0.94), respectively, compared to RMSE of 0.17 (SSIM of 0.93) for supervised training. The independent noise assumption in Noise2Self and Noisier2Noise with image domain noise injection resulted in significantly diminished performance, yielding RMSE of 0.23 and 0.37 (SSIM of 0.86 and 0.84). The NPS measurements revealed a shift towards low frequency components for Noise2Sim, arising from blurring of tissue boundaries and residual image transfer induced by the masking of dissimilar regions in the loss function. Noisier2Noise showed a frequency distribution of noise closer to the high-dose reference. Such performance was slightly degraded for non-matched noise injection models inducing shorter correlation kernels than the nominal detector noise correlation, but models inducing longer correlation showed negligible impact in the denoising results. Self-supervised denoising in presence of correlated noise was proved feasible. Among the evaluated models, Noisier2Noise strategies with projection domain noise injection showed denoising performance comparable to supervised training and noise spectral distribution comparable to high-dose bCT.

In conventional tomosynthesis, the x-ray source or detector move relative to the patient so that anatomy at a target depth is focused and other anatomy is blurred. We propose a real-time single frame tomosynthesis design using a distributed source array and a large flat-panel detector. Each element in the source array energizes simultaneously, and the beam is collimated down so that it passes through isocenter and is received in a small sector of the detector. The detector receives multiple non-overlapping x-ray images simultaneously, and averages these to blur anatomy outside the target depth. Reconstruction occurs at the readout rate of the detector, typically 30 frames per second. Single frame tomosynthesis therefore increases temporal resolution at the expense of field of view and number of views. An application of single frame tomosynthesis is the monitoring of lung tumors during stereotactic body radiotherapy (SBRT). External biplane fluoroscopic systems, presently used for management of cranial lesions, could be repurposed with tomosynthesis at moderate cost. In a reader study with two radiation oncologists evaluating 60 simulated cases of lung SBRT, 90% were deemed acceptable for motion management with tomosynthesis compared to 53% with fluoroscopy. We constructed a prototype system using four portable x-ray sources and a fixed collimator and frame and imaged an anthropomorphic lung phantom with a spherical lung nodule embedded, and found that the prototype system showed displayed the lung nodule with better contrast than fluoroscopy.

Stereotactic ablative radiotherapy (SABR) delivers a high dose of radiation to a small area and is frequently used to treat cancer patients with metastatic lesions in the lung. Selecting the appropriate prescription is a balance of delivering enough radiation to lesions to prevent recurrence, and minimizing the radiation delivered to organs at risk (OARs) to limit side effects. After a radiation oncologist (RO) selects a prescription, treatment planning software is used to create the dose distribution, and calculate radiation delivered to the lesions and OARs. If dose constraints are not met, a different prescription must be selected, and the process is repeated. Planning SABR treatments is resource intensive, and repeated iterations can lead to treatment delays. Recently, machine learning techniques have been used to create a dose distribution for a given SABR prescription. Thus far, these techniques only target single lesions and are not commonly implemented clinically. In this work, we create a conditional generative adversarial network (GAN) with a U-NET backbone to estimate the dose distribution of SABR treatments to 2-6 lesions in the lung. The GAN is conditioned on contours of the OARs and lesions, CT images, and an initial dose estimation. A novel loss function is used during training. Through the mean squared error and dose metrics used by ROs, the output of the GAN demonstrates good agreement with the ground truth dose. The model will allow ROs to efficiently compare prescriptions options, reduce departmental workload by the multidisciplinary team, and circumvent treatment delivery delays for patients.

Histotripsy is an emerging noninvasive, nonthermal, and nonionizing tumor treatment, which uses focused ultrasound to mechanically destroy tissue. Currently, targeting is performed using ultrasound imaging, which can be operator dependent. Additionally, many tumors cannot be seen or are poorly visualized on ultrasound due to obstructions, patient size, location, or echogenicity. In these cases, energy delivery for histotripsy treatment may still be possible if alternative targeting techniques are available. In this study, a mobile C-arm based targeting approach is presented, where a tumor can be selected on a 3D cone beam CT (CBCT) and automatically targeted using a histotripsy system with a robotic arm. To this end, a hand-eye calibration technique between a mobile C-arm and robotic arm was developed, where a CBCT and eight 2D x-ray acquisitions of a calibration phantom in different poses were acquired. The phantom poses in the CBCT coordinate system were estimated using a 2D/3D pose estimation approach and the corresponding robot end effector poses in the robot coordinate system were determined from the joint angles. A dual quaternion approach was then used to determine the relationship between C-arm and robot. During treatment, the histotripsy transducer can be aligned on a target using the calibrated robotic arm. Nine dual-modality soft-tissue mimicking phantoms were treated, and the resulting treatment zones were manually segmented from post-treatment CBCTs as reference. The average Euclidean distance between planned and actual treatment centers was 0.81 ± 0.45mm. This approach could considerably increase the number of patients that could benefit from histotripsy treatment, reduce operator dependency, and facilitate clinical translation of histotripsy.

Percutaneous ablation procedures have been increasingly utilized to non-invasively treat tumors, such as hepatocellular carcinoma, by heating tumor cells beyond the lethal threshold. Intraprocedural temperature monitoring via spectral CT thermometry with a sensitivity less than 3 °C can reduce local recurrence rates by ensuring the tumor and its surrounding safety margin reach lethal temperatures. Because temperature sensitivity is reliant on noise, the effect of additional denoising, radiation dose, slice thickness, and iterative reconstruction levels on temperature sensitivity was evaluated on physical density slices utilized to generate temperature maps. Three different denoising algorithms (total variation, bilateral filtering, and non-local means) were applied to input images prior to generating physical density maps. Differences in noise in physical density and temperature sensitivity were calculated for each combination of parameters. All three denoising algorithms did not significantly affect quantification with an average difference of 1 x 10^-4 g/mL from standard reconstructions, while generally non-local means denoising performed best with noise decreasing to 2 x 10^-4 g/mL. The reduction in noise corresponded to temperature sensitivity decreasing from 15 ± 4 °C with standard reconstructions to 3 ± 2 °C with non-local means denoising at 2 mGy with 2 mm slices. Overall, temperature sensitivity at low radiation doses improved to clinically satisfactory levels with additional denoising. These accurate temperature maps from spectral CT thermometry will enable real-time, non-invasive temperature monitoring to ensure critical structures are not thermally damaged and the entire tumor and safety margin reach the lethal threshold, reducing local recurrences.

Ultrasound computed tomography (USCT) is an emerging imaging modality that holds great promise for breast imaging. Full-waveform inversion (FWI)-based image reconstruction methods incorporate accurate wave physics to produce high spatial resolution quantitative images of speed of sound or other acoustic properties of the breast tissues. However, FWI reconstruction is computationally expensive, which limits its application in a clinical setting. This contribution investigates using the use of a convolutional neural network (CNN) to learn a mapping from USCT data to speed of sound estimates. The CNN was trained using a supervised approach that employed a large set of anatomically and physiologically realistic numerical breast phantoms (NBPs) and simulated USCT measurements. Once trained, this CNN can then be evaluated for real-time FWI image reconstruction from USCT data. The performance of the proposed method was assessed and compared against FWI using a hold-out sample of 41 NBPs and corresponding USCT images. Accuracy was measured using relative mean square error (RMSE) and structural self-similarity index measure (SSIM). This numerical experiment demonstrates that a supervised learning model can achieve accuracy comparable to FWI, while significantly reducing computational time and memory requirements.

Previous studies of dynamic photoacoustic computed tomography (PACT) consider the case where complete data can be rapidly acquired and employed to directly reconstruct a sequence of images. However, such frame-by-frame methods do not apply to commercially available volumetric PACT imaging systems with rotating gantries because the object varies during data acquisition. Furthermore, the rotation speed and the laser repetition rate limit the number of tomographic views per frame. In this study, a low-rank matrix estimation-based spatiotemporal image reconstruction method attuned to rotating-gantry volumetric PACT systems is proposed, and its accuracy is shown by numerical and experimental studies.

Electron paramagnetic resonance imaging (EPRI) is a rising technique for preclinical imaging of small animals. The technique uses paramagnetic spin contrast materials to determine the spectral-spatial (SS) distribution of materials within the subject. A widely used EPRI modality employs continuous wave (CW) scanning scheme with Zeeman modulation (ZM). The imaging model in this technique can be related to the Radon transform (RT) of the SS image, and image reconstruction is equivalent to reconstruction from RT data. However, data collection is limited by the finite strength of the magnetic field gradient applied to the subject, and there is a desire to speed up scanning by collecting data only over a limited-angular range (LAR). In this study, we tailor a recently developed DTV algorithm in CT to investigate accurate image reconstruction from RT over LARs in EPRI. The results show that the DTV algorithm can be adapted for image reconstruction of quality comparable to that of images reconstructed from full-angular range (FAR) data, suggesting that algorithms can be developed to enable LAR scanning in CW-ZM EPRI with reduced imaging time.

Superior material discrimination provided by photon counting detector (PCD) technology promises to transform X-ray CT into a functional and molecular imaging modality while maintaining its high spatial resolution, fast scanning times, and relatively low cost. Our group has developed pre-clinical photon-counting CT (PCCT) prototype systems and applied them, in combination with nanoparticle contrast agents, for cancer and cardiac imaging. This work aims to compare the PCCT imaging performance using a gallium arsenide (GaAs) and a cadmium telluride (CdTe) based PCD, both with 150- μm pixels and 4 energy thresholds. The two PCDs were integrated in the same PCCT system. Phantoms containing elemental solutions of Iodine, Gadolinium, Tantalum, Hafnium, Bismuth and Calcium were imaged with each detector to establish the spectral separation capabilities for PCCT. Moreover, combined dual detectors PCCT imaging was also tested. A joint iterative reconstruction followed by image-based material decomposition was used to provide material maps of different elements. The accuracy of the estimated concentrations within the material decompositions were compared. Our results have shown that GaAs-based PCCT imaging has an overall higher sensitivity (by ~15%) to Iodine than CdTe when using identical acquisition parameters. The CdTe imaging, has higher quantum efficiency at high keVs, supporting higher source kVps and energy threshold settings for imaging the K-edges of Bismuth or other high-Z NPs (such as Gold). Based on the average vial measurements, hybrid GaAs-CdTe PCCT decompositions have the most accurate decompositions for nearly all materials except for Bismuth. The combination of CdTe and GaAs PCDs into a dual source PCCT system will provide high sensitivity in separating multi-element contrast agents from intrinsic tissues.

High spatial resolution CT has enabled a variety of clinical applications. However, further improvements resolution are hampered by current x-ray tubes. Smaller focal spots contribute less focal spot blur but have limited fluence - leading to high levels of noise in CT reconstructions. In this work, we seek to combine multiple focal spots in a single data acquisition to provide both the high-resolution data required to resolve fine features but also the fluence required to reduce noise levels. We demonstrate this technique by using a protocol with two different focal spot sizes to produce multi-resolution data that is jointly processed using a model-based iterative reconstruction (MBIR) method. This novel acquisition and reconstruction scheme is compared with alternate acquisition and processing method (including large focal spot only, and small focal spot only protocols). The multiple focal spot technique exhibits superior performance over a wide range of MBIR regularization strengths - demonstrating the potential of the multi-focal spot approach to surpass traditional high-resolution performance limits.

Tomographic systems based on stationary arrangements of compact x-ray sources coupled to curved panel detectors have shown great potential for point-of-care brain imaging, but suffer from large, non-isotropic x-ray scatter. This work presents an adaptive kernel strategy to efficiently estimate scatter in stationary multi-source CT. The adaptive scatter estimation handles non-circular geometries, by the addition of pre- and post-processing steps to projection domain scatter estimators. The method was calibrated and evaluated on simulated data for a previously presented system with 31 x-ray sources on a circular arc coupled to a curved detector. Further assessment was obtained on experimental data obtained with an imaging testbench including a compact CNT-based x-ray source and simulating the scanner geometry. The method achieved accurate air-normalized scatter distributions across x-ray source positions and detector pixels, yielding a mean absolute error of 1.98𝑥10−3 with respect to the Monte-Carlo ground truth. Air-gap compensation had the largest impact on final accuracy. Image quality for simulated data showed consistent mitigation of scatter artifacts and reduction in non-uniformity from NU = 109 HU to 24 HU, with comparable performance for variations in cranium size, ranging in length from 161 mm (NU =14 HU) to 246 mm (NU = 15 HU). The experimental data showed comparable performance with error attributable to slight simulation infidelity. This work presents an adaptive approach to scatter compensation in multi-source, non-circular geometries using warping and weighting operations coupled to kernel-based scatter estimation on a virtual circular geometry, with immediate extension to other projection-based scatter compensation strategies.

This study evaluated a multi-source cone beam computed tomography (ms-CBCT) design. The design replaces the single X-ray tube in the conventional CBCT with a linear array of collimated X-ray sources aligned along the axial direction. A single beam carbon nanotube (CNT) X-ray source with an adjustable external collimator was mechanically translated to different positions along the axial direction to simulate the configurations with different number of sources and corresponding beam cone angle. A flat panel detector was offset horizontally to extend the field of view. Several phantoms were scanned under different configurations for comparison. The scatter-primary-ratio was reduced essentially linearly with the increasing number of X-ray sources (𝑁_{𝑠𝑜𝑢𝑟𝑐𝑒}) from 47% in the N1 (𝑁_{𝑠𝑜𝑢𝑟𝑐𝑒} = 1, cone angle 𝜃 = 10.3°) to below 20% in the N8 (𝑁_{𝑠𝑜𝑢𝑟𝑐𝑒} = 8, cone angle 𝜃 = 2.3°) configuration. The CT cupping artifact was reduced from 15% in N1 to 10% in N8. The nonuniformity of the CT numbers, as measured by the standard deviation of the Hounsfield Unit (HU) values from multiple regions of interest (ROIs) in the axial plane of a contrast phantom was reduced from 38.0 in the N1 to 19.8 in the N8 configuration. An anthropomorphic head phantom imaged with the N8 ms-CBCT configuration also has a more uniform line profile than the image with the N1 configuration in the central homogenous region.

X-ray spectral imaging has been increasingly investigated in radiography and interventional imaging. Flat-panel detectors with more than one detection layer have demonstrated advantages in providing separate soft tissue and bone images. Current dual-layer flat-panel detectors (DL-FPD) have limited capability to further differentiate between iodinated contrast agent and bony/calcified structures. In this work, we investigate a triple-layer flat-panel detector (TL-FPD) and the feasibility of three-material (water/calcium/iodine) decomposition. A physical model of TL-FPD, including system geometry, spectrum sensitivities, blur and noise models was developed. Using simulated triple-layer projections, three-material decompositions were performed using three different processing methods: polynomial-based, model-based, and a machine learning-based method (ResUnet). We find that the polynomial-based method leads to very noisy images with poor differentiation between calcium and iodine maps. The model-based method achieved much lower noise levels than the polynomial- based method but exhibited residual errors between the iodine and calcium channels. The ResUnet method offered the best decompositions among the investigated methods in terms of root mean square error from the ground truth and noise in the material maps. These preliminary results demonstrate the feasibility of three-material decomposition using TL-FPD and suggest a path for clinical translation of single-shot contrast/iodine differentiation in radiography and fluoroscopy.

Digital subtraction angiography (DSA) is a widely used technique for the visualization of contrast-enhanced structures. However, temporal subtraction DSA is challenged by misregistration artifacts due to patient motion and incomplete separation of iodine contrast agent from background soft tissue and bone. In this work, we propose an approach that allows three-material decomposition using a dual-layer flat panel detector in the presence of soft tissue motion. We assume the calcium signal (bone) remains stationary in the pre- and post-contrast images but allow soft tissues to move freely (e.g. cardiac motion). The dual-layer pre- and post-injection images form and ensemble of four measurements that permits material decomposition of four bases (pre- and post-injection soft tissue, calcium, and iodine). We apply two different processing techniques: 1) a modified lookup table and; 2) a model-based material estimation. These are compared with previously proposed DSA methods using temporal subtraction and hybrid (dual-energy) subtraction. Investigations were performed using an XCAT thorax phantom simulating a breath-hold. The pre- and post-contrast measurements were simulated at different time points within a cardiac cycle. Both the lookup table and model-based algorithms eliminate motion artifact as a result of soft tissue motion and allow good separation of iodine, bone, and soft tissue. While the lookup table algorithm contains high noise at the simulated dose level, the model-based algorithm produced iodine images that allow the visualization of major vessels around the heart. In contrast, traditional temporal DSA is susceptible to subtraction artifacts and hybrid DSA shows increased noise.

Purpose: Physics-informed neural networks (PINNs) and computational fluid dynamics (CFD) have both demonstrated an ability to derive accurate hemodynamics if boundary conditions (BCs) are known. Unfortunately, patient-specific BCs are often unknown, and assumptions based upon previous investigations are used instead. High speed angiography (HSA) may allow extraction of these BCs due to the high temporal fidelity of the modality. We propose to investigate whether PINNs using convection and Navier-Stokes equations with BCs derived from HSA data may allow for extraction of accurate hemodynamics in the vasculature.

Materials and Methods: Imaging data generated from in vitro 1000 fps HSA, as well as simulated 1000 fps angiograms generated using CFD were utilized for this study. Calculations were performed on a 3D lattice comprised of 2D projections temporally stacked over the angiographic sequence. A PINN based on an objective function comprised of the Navier-Stokes equation, the convection equation, and angiography-based BCs was used for estimation of velocity, pressure and contrast flow at every point in the lattice.

Results: Imaging-based PINNs show an ability to capture such hemodynamic phenomena as vortices in aneurysms and regions of rapid transience, such as outlet vessel blood flow within a carotid artery bifurcation phantom. These networks work best with small solution spaces and high temporal resolution of the input angiographic data, meaning HSA image sequences represent an ideal medium for such solution spaces.

Conclusions: The study shows the feasibility of obtaining patient-specific velocity and pressure fields using an assumption-free data driven approach based purely on governing physical equations and imaging data.

Previously we developed a single-shot quantitative x-ray imaging (SSQI) method to perform material decomposition in x-ray imaging by combining the use of a primary modulator (PM) and dual-layer (DL) detector, where the PM removes scatter from DL images while the DL provides dual energy (DE) images for material decomposition (MD), which further removes beam hardening resulting from the partially attenuated regions of the PM. We have demonstrated the concept of SSQI using simulation and further tested its efficacy on chest phantom studies performed on our tabletop system.

In this work, we further explored the clinical value of SSQI in interventional guidance, specifically focusing on investigating its potential in real-time x-ray image guidance. We integrated the SSQI on a C-arm system and designed two studies to evaluate its performance for iodine quantification in both static and dynamic imaging using anthropomorphic phantoms. Compared to direct MD without scatter correction, the RMSE in material-specific images was reduced by 38-64% with SSQI when compared to ground truth. For the dynamic study, the SSQI estimated iodine mass was in close agreement with the amount from a ground truth acquisition. The results in this work further expand the potential of SSQI for real-time image guidance.

Flat panel detectors (FPDs) with two separate scintillator layers enable dual-energy (DE) 2D and 3D x-ray imaging with perfect temporal registration. A thick (1 mm) Cu layer is permanently sandwiched between the two scintillators to widen the spectral separation. Because the majority of clinical applications of FPDs do not benefit from DE imaging, the existence of the Cu layer during those applications degrades the overall dose efficiency. In this work, a novel dual-layer FPD design with a removable liquid filter design was presented to improve the dose efficiency of dual-layer FPDs when they are used for non-DE imaging applications.

Metal Artifacts remain a problem in Cone-Beam CT (CBCT) imaging, especially reducing clinical value in trauma applications by obscuring the important area around implants. Building on existing Metal Artifact Avoidance (MAA) algorithms, we formulate a metric based on the scatter fraction, introduce a shape model which reduces computational requirements and analyze the feasibility of using only two scout views as prior information.

Driven by the observation that orbit optimization requires knowledge of the position, extent, and orientation of metallic objects in the volume of interest (VOI), we devise a shape model in the form of ellipsoids. Reconstructing an ellipsoid from two projection images is not unambiguously possible using analytic methods. By interpreting the problem as probability density estimation, a maximum likelihood fit can be recovered using a Gaussian Mixture Model (GMM). This parametric representation of metal objects is used to efficiently calculate metal pathlength maps for candidate projections on tilted circular trajectories through analytic forward projection. The scatter fraction behind the metal object is modelled as a function of metal pathlength to score views and choose artifact minimizing tilted trajectories.

Given two projection images of simulated ellipsoidal objects, the GMM accurately estimates the position and longest axis length within millimeter tolerance. Depending on the orientation relative to the two acquired scout views, an estimation error within the scout-view plane is observed. The generalization on measured data and the shape model hypothesis is verified in a phantom study showing a good correspondence of the modelled metric and observed reduction of artifacts in the tilted CBCT scans.

The severity of Metal Artifacts can be reduced by optimizing trajectories on low-fidelity shape models. These surrogate representations can be efficiently estimated from two views for most relative object orientations, but a well-defined ‘blind spot’ remains. The reconstruction error was found to have little effect on tilted orbit optimization if the tilt-axis is contained in the scout-view plane.

To compare the image quality characteristics of a cone-beam CT (CBCT) imaging system with noncircular scan protocols (Sine-on-Sphere) to a conventional circular orbit. A phantom was used to assess image uniformity, image noise, and noise-power spectrum (NPS) for a biplane C-arm (Icono; Siemens Healthineers) that could perform CBCT with either a circular or non-circular orbit (Sine-on-Sphere, “SineSpin”). The non-circular orbit used a sinusoidal motion with an amplitude of ±10º during the orbit. Uniformity and quantum noise were similar for the circular and non-circular orbits. A slight increase in cupping artifact for the noncircular orbit was observed and attributed to slightly increased path lengths (increased scatter-to-primary ratio) for oblique rays through the cylindrical phantom. The NPS clearly depicted closing of the “null cone” about the fz spatial frequency axis owing to improved sampling with the non-circular orbit. A non-circular orbit for CBCT resulted in improved sampling with similar noise magnitude and uniformity compared to a circular orbit.

Cone-beam CT (CBCT) is widespread in abdominal interventional imaging, but its long acquisition time makes it susceptible to patient motion. Image-based autofocus has shown success in CBCT deformable motion compensation, via deep autofocus metrics and multi-region optimization, but it is challenged by the large parameter dimensionality required to capture intricate motion trajectories. This work leverages the differentiable nature of deep autofocus metrics to build a novel optimization strategy, Multi-Stage Adaptive Spine Autofocus (MASA), for compensation of complex deformable motion in abdominal CBCT. MASA poses the autofocus problem as a multi-stage adaptive sampling strategy of the motion trajectory, sampled with Hermite spline basis with variable amplitude and knot temporal positioning. The adaptive method permits simultaneous optimization of the sampling phase, local temporal sampling density, and time-dependent amplitude of the motion trajectory. The optimization is performed in a multi-stage schedule with increasing number of knots that progressively accommodates complex trajectories in late stages, preconditioned by coarser components from early stages, and with minimal increase in dimensionality. MASA was evaluated in controlled simulation experiments with two types of motion trajectories: i) combinations of slow drifts with sudden jerk (sigmoid) motion; and ii) combinations of periodic motion sources of varying frequency into multi-frequency trajectories. Further validation was obtained in clinical data from liver CBCT featuring motion of contrast-enhanced vessels, and soft-tissue structures. The adaptive sampling strategy provided successful motion compensation in sigmoid trajectories, compared to fixed sampling strategies (mean SSIM increase of 0.026 compared to 0.011). Inspection of the estimated motion showed the capability of MASA to automatically allocate larger sampling density to parts of the scan timeline featuring sudden motion, effectively accommodating complex motion without increasing the problem dimension. Experiments on multifrequency trajectories with 3-stage MASA (5, 10, and 15 knots) yielded a twofold SSIM increase compared to single-stage autofocus with 15 knots (0.076 vs 0.040, respectively). Application of MASA to clinical datasets resulted in simultaneous improvement on the delineation of both contrast-enhanced vessels and soft-tissue structures in the liver. A new autofocus framework, MASA, was developed including a novel multi-stage technique for adaptive temporal sampling of the motion trajectory in combination with fully differentiable deep autofocus metrics. This novel adaptive sampling approach is a crucial step for application of deformable motion compensation to complex temporal motion trajectories.

Purpose: We investigated the feasibility of dual-energy (DE) detection of bone marrow edema (BME) using a dedicated extremity cone-beam CT (CBCT) with a unique three-source x-ray unit. The sources can be operated at different energies to enable single-scan DE acquisitions. However, they are arranged parallel to the axis of rotation, resulting in incomplete sampling and precluding the application of DE projection-domain decompositions (PDD) for beam-hardening reduction. Therefore, we propose a novel combination of a model-based “one-step” DE two-material decomposition followed by a constrained image-domain change-of-basis to obtain virtual non-calcium (VNCa) images for BME detection.

Methods: DE projections were obtained using an “alternating-kV” protocol by operating the peripheral two sources of the CBCT system at low-energy (60 kV, 0.105 mAs/frame) and the central source at high-energy (100 kV, 0.028 mAs/frame), for a total of 600 frames over 216° of gantry rotation. Projections were processed with detector lag, glare and fast Monte Carlo (MC)-based iterative scatter corrections. Model-based material decomposition (MBMD) was then implemented to obtain aluminum (Al) and polyethylene (PE) volume fraction images with minimal beam-hardening. Statistical ray weights in MBMD were modified to account for regions with highly oblique sampling by the peripheral sources. To generate the VNCa maps, image-domain decomposition (IDD) constrained by the volume conservation principle (VCP) was performed to convert the Al and PE MBMD images into volume fractions of water, fat and cortical bone. Accuracy of BME detection was evaluated using physical phantom data acquired on the multi-source extremity CBCT scanner.

Results: The proposed framework estimated the volume of BME with ~10% error. The MC-based scatter corrections and the modified MBMD ray weights were essential to achieve such performance – the error without MC scatter corrections was <30%, whereas the uniformity of estimated VNCa images was 3x improved using the modified weights compared to the conventional weights.

Conclusions: The proposed DE decomposition framework was able to overcome challenges of high scatter and incomplete sampling to achieve BME detection on a CBCT system with axially-distributed x-ray sources.

This conference presentation was prepared for the Medical Imaging 2023: Physics of Medical Imaging conference at SPIE Medical Imaging 2023.

Cone-beam computed tomography (CBCT) is widely used in patient positioning for image-guided radiotherapy. However, compared to simulation CT, it suffers from scattering and beam hardening artifacts, Hounsfield units (HU) value inaccuracy, low tissue contrast, and high noise, which impede its potential applications, such as adaptive radiotherapy through online contouring and dose calculation. The capabilities of dual-energy CBCT (DE-CBCT) to overcome the aforementioned issues has been demonstrated in preclinical studies by introducing extra information along the spectral dimension and enabling the application of material decomposition. In this work, we propose a CT guided and sparsity constrained multi-material decomposition method for DE-CBCT to improve the image quality and quantitative accuracy of both DE-CBCT and material composition images. First, a local linear constraint is employed to mathematically describe the structure similarities between the DE-CBCT/ basis material and CT images, so that CT can effectively work as a guidance by introducing its superior image quality. Secondly, to regulate the multi-material decomposition process, low rank approximation with trace norm truncation is employed to limit the number of material bases and incorporate sum-to-one constraint to bound the material volume fractions. In addition, Mumford-Shah regularization in spatial domain is introduced for edge preservation and piece-wise smooth. Both phantom and patient studies demonstrate the superiority of the proposed method in image-quality, decomposition-accuracy, edge preservation, noise suppression and artifacts reduction, which further benefits some potential applications, such as online contouring and dose calculation for adaptive radiotherapy.

Contrast-enhanced digital breast tomosynthesis (CEDBT) utilizes iodinated contrast agents and spectral imaging techniques to highlight breast tumors with neo-angiogenesis. Conventional dual-shot (DS) CEDBT suffers from motion artifacts due to patient motion between two separate x-ray exposures for the sequential acquisitions of high- and low-energy (HE and LE) images. To mitigate motion artifacts, a new CEDBT system was designed with a direct-indirect dual-layer (DL) detector. It enables simultaneous acquisition of HE and LE images with a single exposure. Our previous study has demonstrated the feasibility of the DL detector to cancel normal breast tissue and provide lesion contrast comparable to DS. Due to the absence of anti-scatter grids, both DS- and DL-CEDBT suffer image quality degradation from scattered radiation, which introduces cupping artifacts and reduces lesion contrast. In this study, we are proposing a hybrid scatter correction method, which combines the straightforward kernel-based iterative convolution method and the empirical interpolation method. The hybrid method was applied to Monte Carlo (MC) simulated DL-DBT projection images of anthropomorphic digital breast phantoms of varying glandularity and compressed thickness. Results showed that the hybrid method improves scatter estimation accuracy compared to the single-kernel iterative convolution method. The hybrid method was applied to a clinical DS-CEDBT case for scatter correction. The results showed reduced cupping artifacts and improved lesion contrast.

Image noise in digital breast tomosynthesis (DBT) reduces the detectability of subtle signs of breast cancer such as microcalcifications (MC). This study investigated the potential of applying DNGAN, our previously developed deep convolutional neural network (CNN) DBT denoiser, to different reconstruction stages to improve the image quality, including projection views before reconstruction, intermediate images during iterative reconstruction, or final reconstructed images, and a combination of the different stages of denoising. We also proposed two CNNs as task-based image quality measures to compare different reconstructions: a CNN noise estimator (CNN-NE) trained to evaluate the noise level of a given DBT image, and a CNN MC classifier (CNN-MC) trained to estimate the detectability of MCs by classifying clustered MCs from MC-free backgrounds. The CNN-NE was trained with virtual DBTs reconstructed from projections generated by the VICTRE tool over a wide range of noise levels. The CNN-MC was trained with human subject DBTs. We adopted the training strategy of transfer learning to train CNN-NE and CNN-MC due to the limited training data. We found that the increase in AUC estimated by the CNN-MC classifier correlated well with the decrease in image noise by DNGAN estimated by CNN-NE on an independent human subject test set. A combination of DNGAN-regularized plug-and-play reconstruction and an additional DNGAN post-reconstruction denoising achieved the lowest noise level and the best MC detectability. The AUC and noise rankings from the CNNs matched our visual judgement that less noisy images had better MC conspicuity.

We describe a longitudinal in silico imaging trial investigating the advantages of digital breast tomosynthesis (DBT) versus digital mammography (DM) for early detection of breast cancer. To mimic cancer progression, we used a previously developed computational model based on biological and physiological phenomena accounting for rate of metabolic nutrients and cellular waste as well as the stiffness of surrounding anatomical structures affecting lesion morphology. We integrated this model with the VICTRE pipeline to create a cohort of in silico patients each with a unique manifestation of cancer recorded at 5 stages of progression. Digital patients with varying breast densities were considered. A customized version of the VICTRE pipeline was used to simulate DM and DBT imaging of patients with an in silico version of the Siemens Mammomat Inspiration system with image interpretation under a location-known-exactly tasks, relying on 2D/3D algorithmic readers previously described. We analyzed the area under the ROC curve (AUC) for both imaging modalities at the 5 stages of cancer growth to evaluate the performance of DBT and DM along the life of the tumor. Our findings suggest that DBT outperforms DM for all lesion sizes, which is consistent with studies reported in literature. We observed the mean AUCs increases from 0.64 to 0.80 (p < 0.001) forDMand from 0.66 to 0.88 (p < 0.001) for DBT as lesion size increased from 0.37 to 1.8 mm. These results suggest a potential benefit of DBT as compared to DM for the detection of small masses at earlier stages of tumor development. The in silico trial we designed allowed for studying the progression of detectability of masses at different growing stages, something that would be costly and ethically questionable with a human clinical trial.

Chronic obstructive pulmonary disease (COPD) is one of the top three causes of death worldwide, characterized by emphysema and bronchitis. Airway measurements reflect the severity of bronchitis and other airway-related diseases. Airway structures can be objectively evaluated with quantitative computed tomography (CT). The accuracy of such quantifications is limited by the spatial resolution and image noise characteristics of the imaging system and can be potentially improved with the emerging photon-counting CT (PCCT) technology. This study evaluated the quantitative performance of PCCT against energy-integrating CT (EICT) systems for airway measurements, and further identified optimum CT imaging parameters for such quantifications. The study was performed using a novel virtual imaging framework by developing the first library of virtual patients with bronchitis. These virtual patients were developed based on CT images of confirmed COPD patients with varied bronchitis severity. The human models were virtually imaged at 6.3 and 12.6 mGy dose levels using a scanner-specific simulator (DukeSim), synthesizing clinical PCCT and EICT scanners (NAEOTOM Alpha, FLASH, Siemens). The projections were reconstructed with two algorithms and kernels at different matrix sizes and slice thicknesses. The CT images were used to quantify clinically relevant airway measurements (“Pi10” and “WA%”) and compared against their ground truth values. Compared to EICT, PCCT provided more accurate Pi10 and WA% measurements by 63.1% and 68.2%, respectively. For both technologies, sharper kernels and larger matrix sizes led to more reliable bronchitis quantifications. This study highlights the potential advantages of PCCT against EICT in characterizing bronchitis utilizing a virtual imaging platform.

This work proposes an objective and automated procedure to obtain realistic digital breast tomosynthesis (DBT) images for virtual clinical trials (VCT) using the hybrid approach (simulating lesions in acquired patient images). Based on extensive feedback from a radiologist, we have implemented an automatic selection of an appropriate insertion position that (1) is located in the interior region of the breast, (2) contains sufficient glandular tissue, and (3) has the lowest variance to cope with the presence of prominent background structure and contains no Sobel edge detected blood vessels. Next, the lesion is rotated to align with the breast structures using the histogram of oriented gradient feature descriptor. The spicules of the lesion are extended to improve the fine details of the manually segmented mass models. To reduce the pronounced shadow artefact surrounding the mass, the lesion template is modified with a fitted 2D gaussian to create a softer transition between background and lesion. The realism of 20 simulated lesions using the established automated procedure was scored; 70% of the simulated cases received at least a realism score of 4 out of 5. This means an improvement in realism compared to when lesions without processing are inserted at random locations in the patient background images. Additionally, the automated method eliminates the dependency on the researcher performing the VCT.

The performance of a CT scanner for detectability tasks is difficult to precisely measure. Metrics such as contrast-to-noise ratio, modulation transfer function, and noise power spectrum do not predict detectability in the context of nonlinear reconstruction. We propose to measure detectability using a dense search challenge: a phantom is embedded with hundreds of target objects at random locations, and a human or numerical observer analyzes the reconstruction and reports on suspected locations of all target objects. The reported locations are compared to ground truth to produce a figure of merit, such as area under the curve (AUC), that is sensitive to the acquisition dose and the dose efficiency of the CT scanner. We used simulations to design such a dense search challenge phantom and found that detectability could be measured with precision better than 5%. Test 3D prints using the PixelPrint technique showed the feasibility of this technique.

Attenuation compensation (AC) is beneficial for visual interpretation tasks in single-photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI). However, traditional AC methods require the availability of a transmission scan, most often a CT scan. This approach has the disadvantage of increased radiation dose, increased scanner costs, and the possibility of inaccurate diagnosis in cases of misregistration between the SPECT and CT images. Further, many SPECT systems do not include a CT component. To address these issues, we developed a Scatter-window projection and deep Learning-based AC (SLAC) method to perform AC without a separate transmission scan. To investigate the clinical efficacy of this method, we then objectively evaluated the performance of this method on the clinical task of detecting perfusion defects on MPI in a retrospective study with anonymized clinical SPECT/CT stress MPI images. The proposed method was compared with CT-based AC (CTAC) and no-AC (NAC) methods. Our results showed that the SLAC method yielded an almost overlapping receiver operating characteristic (ROC) plot and a similar area under the ROC (AUC) to the CTAC method on this task. These results demonstrate the capability of the SLAC method for transmission-less AC in SPECT and motivate further clinical evaluation.

Dual-energy computed tomography (DECT) enables material decomposition for tissues and produces additional information for PET/CT imaging to potentially improve the characterization of diseases. PET-enabled DECT (PDECT) allows the generation of PET and DECT images simultaneously with a conventional PET/CT scanner without the need for a second x-ray CT scan. In PDECT, high-energy γ-ray CT (GCT) images at 511 keV are obtained from time-of-flight (TOF) PET data and are combined with the existing x-ray CT images to form DECT imaging. We have developed a kernel-based maximum-likelihood attenuation and activity (MLAA) method that uses x-ray CT images as a priori information for noise suppression. However, our previous studies focused on GCT image reconstruction at the PET image resolution which is coarser than the image resolution of the x-ray CT. In this work, we explored the feasibility of generating super-resolution GCT images at the corresponding CT resolution. The study was conducted using both phantom and patient scans acquired with the uEXPLORER total-body PET/CT system. GCT images at the PET resolution with a pixel size of 4.0 mm × 4.0 mm and at the CT resolution with a pixel size of 1.2 mm × 1.2 mm were reconstructed using both the standard MLAA and kernel MLAA methods. The results indicated that the GCT images at the CT resolution had sharper edges and revealed more structural details compared to the images reconstructed at the PET resolution. Furthermore, images from the kernel MLAA method showed substantially improved image quality compared to those obtained with the standard MLAA method.

Due to various physical degradation factors, the signal-to-noise ratio (SNR) and image quality of PET needs further improvements. In this work, we proposed a denoising diffusion probabilistic model-based framework for PET image denoising, where the MR prior image was supplied as the additional network input, and the PET information was included in the iterative refinement steps based on Gaussian distribution assumption. 140 ¹⁸F-MK- 6240 datasets were used in the evaluation, with 1/4 and 1/8 low-dose levels tested for different methods. Global and regional quantifications show that the proposed framework can outperform the Unet-based denoising and MR-guided nonlocal mean denoising methods.

The recently developed total-body positron emission tomography (PET) scanner can capture total-body tracer kinetics, thus enabling systems imaging of multiple organs and interactions between organs, simultaneously. However, their prohibitive cost and sitting requirements may present a barrier for widespread adaptation. Most commercial PET scanners have a limited AFOV. To circumvent the limited AFOV, scanner manufacturers have implemented static step and shoot (SSS) protocols to “stitch” images acquired at multiple bed positions into a single total-body image. However, the resulting “total-body” images may not be quantitative depending on the kinetics of the tracers, thus biasing quantitative imaging comparisons. We propose a dynamic step and shoot (DSS) protocol with 2sec temporal sampling to pursue a continuous imaging protocol with different acquisition times in different bed positions for each pass through the torso. D-optimal criterion was used to optimize the acquisition protocol using a simulated annealing algorithm. The overall approach is illustrated in estimating parameters of a reversible two-compartment PET kinetic model for key organs in the torso. The intra- and inter-subject performance of the optimal DSS (ODSS) protocol was compared with the SSS protocol in terms of bias and variability and in comparison to the total-body (TB) protocol. The simulations suggest that the proposed ODSS protocol outperforms the conventional SSS protocol in both intra- and inter-subject parameter accuracy and precision tests and generates similar macro-parameter estimates compared to the TB scanner. Overall, we demonstrate that we can achieve an optimal temporal imaging schedule to support quantitative TB systems imaging.

Many imaging systems can be described by a linear operator that maps object properties to a collection of discrete measurements. The null space of such an imaging operator represents the set of object components that are effectively invisible to the imaging system. The ability to extract the object components that lie within the null space of an imaging operator allows one to analyze and optimize not only the measurement capabilities of the system itself, but also its associated reconstruction methods. An orthogonal null space projection operator (ONPO), which maps any object to its corresponding null space component, offers this ability. However, existing methods for producing an ONPO are limited by high memory requirements. In this work, we develop a novel learning-based method for calculating an ONPO. Numerical results show that our method can produce an accurate ONPO using less memory than existing methods, enabling the characterization of the null spaces of larger imaging operators than previously possible.

Deep convolutional neural network (DCNN)-based noise reduction methods have been increasingly deployed in clinical CT. Accurate assessment of their spatial resolution properties is required. Spatial resolution is typically measured on physical phantoms, which may not represent the true performance of DCNN in patients as it is typically trained and tested with patient images and the generalizability of DNN to physical phantoms is questionable. In this work, we proposed a patient-data-based framework to measure the spatial resolution of DCNN methods, which involves lesion- and noise-insertion in projection domain, lesion ensemble averaging, and modulation transfer function measurement using an oversampled edge spread function from the cylindrical lesion signal. The impact of varying lesion contrast, dose levels, and CNN denoising strengths were investigated for a ResNet-based DCNN model trained using patient images. The spatial resolution degradation of DCNN reconstructions becomes more severe as the contrast or radiation dose decreased, or DCNN denoising strength increased. The measured 50%/10% MTF spatial frequencies of DCNN with highest denoising strength were (-500 HU:0.36/0.72 mm^-1; -100 HU:0.32/0.65 mm^-1; -50 HU:0.27/0.53 mm^-1; -20 HU:0.18/0.36 mm^-1; -10 HU:0.15/0.30 mm^-1), while the 50%/10% MTF values of FBP were almost kept constant of 0.38/0.76 mm^-1.

Assessing the reliability of convolutional neural network (CNN)-based CT imaging techniques is critical for reliable deployment in practice. Some evaluation methods exist but require full access to target CNN architecture and training data, something not available for proprietary or commercial algorithms. Moreover, there is a lack of systematic evaluation methods. To address these issues, we propose a patient-specific uncertainty and bias quantification (UNIQ) method that integrates knowledge distillation and Bayesian deep learning. Knowledge distillation creates a transparent CNN (“Student CNN”) to approximate the target non-transparent CNN (“Teacher CNN”). Student CNN is built as a Bayesian-deep-learning-based probabilistic CNN that, for each input, always generates statistical distribution of the corresponding outputs, and characterizes predictive mean and two major uncertainties – data and model uncertainty. UNIQ was evaluated using a low-dose CT denoising task. Patient and phantom scans with routine-dose and synthetic quarter-dose were used to create training, validation, and testing sets. To demonstrate, Unet and Resnet were used as backbones of Teacher CNN and Student CNN respectively and were trained using independent training sets. Student Resnet was qualitatively and quantitatively evaluated. The pixel-wise predictive mean, data uncertainty, and model uncertainty from Student Resnet were very similar to the counterparts from Teacher Unet (mean-absolute-error: predictive mean 1.5HU, data uncertainty 1.8HU, model uncertainty 1.3HU; mean 2D correlation coefficient: total uncertainty 0.90, data uncertainty 0.86, model uncertainty 0.83). The proposed UNIQ can potentially systematically characterize the reliability of non-transparent CNN models used in CT.

Supervised deep learning methods have rapidly advanced to the forefront of medical imaging research; however, they face limitations in advanced CT applications. For instance, label generation is difficult for retrospective cardiac CT data where dose modulation yields variable image quality by phase. Similarly, a model trained to denoise multi-energy data may not generalize across differing contrast and energy channel counts. In this work, we propose and demonstrate several innovations for unsupervised denoising of such multi-channel CT data sets with a focus on improving model robustness to varying noise levels, image contrast, and channel counts. Specifically, we update our past Constrained Bregman Framework for unsupervised denoising to simplify model training procedures, to automate regularization hyperparameter selection, and to spatially adapt denoising performance. Furthermore, we propose a new data normalization strategy involving weighted singular value decomposition, convolutional mean subtraction, and noise variance scaling which abstracts the image denoising problem from the contrast and channel count of the input data. We combine these improvements with a hybrid Inception Net, half U-Net network structure, ideally suited for 3D data processing, and a plug-and- play scheme for incorporating traditional regularizers into deep learning cost functions. We demonstrate the value of these improvements for network training and denoising using retrospectively gated cardiac photon-counting CT data acquired with a Siemens NAEOTOM Alpha scanner (2-3x noise reduction). This clinical model robustly generalizes to 40-channel preclinical cardiac, photon-counting CT data acquired in a mouse without the need for additional training (6- 10x noise reduction, negligible intensity bias).

Low-contrast lesions are difficult to detect in noisy low-dose CT images. Improving CT image quality for this detection task has the potential to improve diagnostic accuracy and patient outcomes. In this work, we use tunable neural networks for CT image restoration with a hyperparameter to control the variance/bias tradeoff. We use clinical images from a super-high-resolution normal-dose CT scan to synthesize low-contrast low-dose CT images for supervised training of deep learning CT reconstruction models. Those models are trained using with multiple noise realizations so that variance and bias can be penalized separately. We use a training loss function with one hyperparameter called the denoising level, which controls the variance/bias tradeoff. Finally, we evaluate the CT image quality to find the optimal denoising level for low-contrast lesion detectability. We evaluate performance using a shallow neural network model classification model to represent a suboptimal image observer. Our results indicate that the optimal networks for low-contrast lesion detectability are those that prioritize bias reduction rather than mean-squared error, which demonstrates the potential clinical benefit of our proposed tunable neural networks.

In recent years there has been increased focus on further reducing radiation dose in CT with photon counting CT using solid-state direct-conversion photon counting detectors (PCDs) to reduce the effective dose from routine CT exams to less than 1 mSv. However, despite its noise-reducing capabilities, PCD-CT faces challenges of inaccurate CT numbers at low-dose levels: with smaller pixel areas and multiple energy channels, the number of digital counts recorded in each bin of each PCD pixel can be as low as single-digit integers leading to statistical biases in CT sinograms due to the nonlinear log transformation operation. After tomographic reconstruction, those biases lead to inaccurate CT numbers in PCD-CT images. Previous correction methods require access to the original raw PCD counts. However, in almost all commercial CT systems, raw detector counts are hidden from the end users. Additionally, some CT systems perform the logarithmic transformation of raw counts as a part of the analog-to-digital conversion process for data compression reasons. For those systems, access to the PCD counts is irretrievably lost. Even for the post-log sinogram data, they are usually not archived for each patient. These practical considerations present challenges to the offline application of CT number bias corrections. The purpose of this work was to develop a method to address the statistical bias problem in low-dose PCD-CT without requiring any access to the raw detector counts. Innovations were made in this work to enable bias correction using the post-log sinogram data or using the reconstructed, bias-contaminated PCD-CT images.

In CT imaging, a standard set of bench testing performances, including modulation transfer function (MTF) and noise power spectrum (NPS), is considered essential to ensure that patient images from CT systems have sufficient image quality. These performances are measured using CT quality assurance (QA) phantoms commonly of a uniform background. However, for deep learning (DL)-based CT image denoising models that are often overwhelmingly trained with patient CT images, it is unclear whether bench testing performances measured with a uniform background reflect the performance in anatomic patient backgrounds. In this work, we insert test objects into a uniform phantom and a patient slice and simulate their CT images to facilitate the measurement of contrast-dependent MTF (cdMTF) and NPS, which are indicators for image sharpness and noise properties. We compare the cdMTF and NPS behaviors of three DL denoising models (REDCNN, REDCNN-tv and DnCNN) measured with images of uniform and patient backgrounds. We observed that cdMTF performance appears consistent between backgrounds, but noise reduction and texture performances can vary substantially. We suggest that caution should be taken when conclusions regarding noise-related performances of a DL-based CT image denoiser are made purely based on measurements from a uniform phantom.

The correction of zero-counts is essential to enable high-quality and accurate low-dose photon counting detector CT (PCD-CT) imaging. Prior to logarithmic normalization, it is necessary to correct for zero-counts in the raw output of PCDs to prevent division-by-zero in the conventional CT reconstruction pipeline. The probability of registering a count of zero for a given detector element is significantly higher for PCDs compared with conventional detectors due to the increased spatial resolution of PCDs and the allocation of counts to multiple energy bins for spectral CT imaging applications. This concern becomes further amplified when considering low-dose PCD-CT applications, large patient imaging, and the presence of metallic structures. However, current methods to correct for zero-counts introduce CT number bias, degrade spatial resolution, or violate the conditional independence assumption. The objective of this work is to develop a zero-correction framework that effectively addresses zero-counts while minimizing bias, preserving conditional independence, and maintaining spatial resolution. Experimental validation studies are performed on a benchtop PCD-CT system to demonstrate the efficacy and generalizability of the proposed correction framework. The results of these studies indicate that the proposed zero-count correction scheme can minimize bias and preserve spatial resolution for both single-energy bin and dual-energy bin PCD-CT imaging acquisitions at low doses.

The rendition of medical images influences the accuracy and precision of quantifications. Image variations or biases make measuring imaging biomarkers challenging. The objective of this paper is to reduce the variability of computed tomography (CT) quantifications for radiomics and biomarkers using physics-based deep neural networks (DNNs). The proposed framework used the knowledge of ground truth through a virtual imaging trial (VIT) methodology to harmonize the different renditions of CT scans across variations in reconstruction kernel and dose. This harmonization was done by developing a generative adversarial network (GAN) model, informed by pixel values and patient-based modulation transfer function (MTF) estimates. To train the network, the VIT platform was used to acquire CT images from a set of forty computational human models (XCAT). Models included varying levels of pulmonary diseases including lung nodules and emphysema. The patient models were imaged with a validated CT simulator (DukeSim) modeling a commercial CT scanner operating across a range of dose (20-100 mAs) and reconstruction kernels (15 from smooth to sharp). The harmonized virtual images were evaluated in three different ways: 1) visual assessment of the images, 2) bias and variation in density-based biomarkers, and 3) bias and variation in morphological-based biomarkers. The harmonization improved the bias and variability of the test set images yielding a structural similarity index of 95±1%, a normalized mean squared error of 10.2±1.5%, and a peak signal-to-noise ratio of 31.8±1.5 dB compared to variations in these metrics when no harmonization was applied (87±9%, 14.2±4%, and 29.6±2, respectively). Emphysema-based imaging biomarkers of LAA-950 (-1.5±1.8%), Perc15 (13.65±9.3 HU), and Lung mass (0.1±0.3 g) had more precise and consistent quantifications compared to values when images were not harmonized (9.9±12%, -36.0±53 HU, and 0.2±0.4g, respectively). These results suggest that the method can be promising to improve consistency in multi-center studies when reliable and consistent quantification of data from multiple systems are required.

Cardiac CT plays an important role in diagnosing heart diseases but is conventionally limited by its complex workflow that requires dedicated phase and bolus tracking [e.g., electrocardiogram (ECG) gating]. This work reports initial progress towards robust and autonomous cardiac CT exams through deep learning (DL) analysis of pulsed-mode projections (PMPs). To this end, cardiac phase and its uncertainty were simultaneously estimated using a novel projection domain cardiac phase estimation network (PhaseNet), which utilizes a sliding-window multi-channel feature extraction approach and a long short-term memory (LSTM) block to extract temporal correlation between time-distributed PMPs. Monte-Carlo dropout layers were utilized to predict the uncertainty of deep learning-based cardiac phase prediction. The performance of the proposed phase estimation pipeline was evaluated using accurate physics-based emulated data.

PhaseNet demonstrated improved phase estimation accuracy compared to more standard methods in terms of RMSE (~43% improvement vs. a standard CNN-LSTM; ~17% improvement vs. a multi-channel residual network [ResNet]), achieving accurate phase estimation with <8% RMSE in cardiac phase (phase ranges from 0-100%). These findings suggest that the cardiac phase can be accurately estimated with the proposed projection domain approach. Combined with our previous work on PMP-based bolus curve estimation, the proposed method could potentially be used to achieve autonomous cardiac CT scanning without ECG device or expert-in-the-loop bolus timing.

Numerous dual-energy CT (DECT) techniques have been developed in the past few decades. Dual-energy CT (DECT) statistical iterative reconstruction (SIR) has demonstrated its potential for reducing noise and increasing accuracy. Our lab proposed a joint statistical DECT algorithm for stopping power estimation and showed that it outperforms competing image-based material-decomposition methods. However, due to its slow convergence and the high computational cost of projections, the elapsed time of 3D DECT SIR is often not clinically acceptable. Therefore, to improve its convergence, we have embedded DECT SIR into a deep learning model-based unrolled network for 3D DECT reconstruction (MB-DECTNet) that can be trained in an end-to-end fashion. This deep learning-based method is trained to learn the shortcuts between the initial conditions and the stationary points of iterative algorithms while preserving the unbiased estimation property of model-based algorithms. MB-DECTNet is formed by stacking multiple update blocks, each of which consists of a data consistency layer (DC) and a spatial mixer layer, where the spatial mixer layer is the shrunken U-Net, and the DC layer is a one-step update of an arbitrary traditional iterative method. Although the proposed network can be combined with numerous iterative DECT algorithms, we demonstrate its performance with the dual-energy alternating minimization (DEAM). The qualitative result shows that MB-DECTNet with DEAM significantly reduces noise while increasing the resolution of the test image. The quantitative result shows that MB-DECTNet has the potential to estimate attenuation coefficients accurately as traditional statistical algorithms but with a much lower computational cost.

PET image reconstruction direct from list-mode data can eliminate the storage of empty sinogram bins, preserve all the precision and accuracy of the large amount data of PET scanners. However, the traditional list-mode reconstruction methods, such as list-mode ML-EM algorithm always suffer from high level of noise due to the ill-conditioning of the PET reconstruction problem. In this paper, we proposed a novel deep learning based method for list-mode reconstruction. We first adopted a domain transfer function to convert the sensor domain data to image domain, then the Residual Attention Dense U-net was used to learn the reconstruction.The proposed RADU-net is based on the RDU-net structure, where the Attention Gate module is integrated into.Instead of concatenating the left side feature to the right side directly, the attention gated module takes advantage of the high level feature to guide the low level feature to realize the attention mechanism before doing the concatenation. Realistic PET acquisitions of 25 digital PET brain phantoms were simulated, generating noisy list-mode data, used for evaluation. Quantification results show that the proposed listmodeCNN can outperform the U-net, list-mode ML-EM as well as TV regularized ML-EM in terms of SSIM and PSNR.

Photon Counting CT (PCCT), which can obtain the energy information of each X-ray photon, is expected to provide new diagnoses. In the conventional CT, iodine contrast agent is often used to evaluate the tumor malignancy, and the capability of PCCT was evaluated about the discrimination between malignant tumors affected by iodine and some objects without iodine like blood clots, which is not necessarily be feasible without the energy information. The phantom with rods containing iohexol (0 - 3 mgI/mL) as the malignant tumor analogue and sucrose (0 - 200 mg/mL) as the benign tumor was used and material decomposition was performed into tin and PMMA as the basis materials. ROI values were derived in the basis material images and 70 keV virtual monochromatic image (VMI) for each rod. As a result, some rods of different compositions showed almost the same CT values in the 70 keV VMI, while the two-dimensional plot of tin and PMMA values demonstrated that the contribution of iodine and sucrose was clearly separated. This would mean that PCCT may bring improvement of quantitative iodine evaluation and better diagnosis on malignant and benign tumors.

Deep-learning (DL) is used extensively in the noise reduction of x-ray computed tomography. In DL, the availability of ground-truth images (GTI) is critical. GTI contains all anatomical information of the low-dose dataset at a much-reduced noise level, resembling that of a high-dose scan of the object. Clinical GTI, however, is difficult to obtain because of ethical and practical constraints. Although the use of model-based-iterative-reconstruction (MBIR) has been proposed, resulting DL images inherent undesired noise characteristics of MBIR.

We propose an approach to utilize low-dose FBP images to generate GTI. The approach is based on the observation that the noise in a thick-slice image (TSI) is suppressed while good texture is maintained. Unfortunately, anatomical details of TSI are degraded, resulting from z-averaging. The proposed approach takes the subtraction-image (SI) between the original and TSI, and performs iterative noise reduction (guided by the standard deviation of SI) to gradually remove noise in SI. After the processing, only the differential anatomical structure between the original and TSI remains. This structure is then subtracted from TSI to arrive at GTI. The amount of noise reduction can be adjusted by modifying the slice-thickness of TSI (a factor of 2.4 noise reduction is selected for this study). This approach is evaluated extensively with both clinical neural and body images. For robustness evaluation, all parameters were kept unchanged. For better visual inspection, difference images between the original and GTI are evaluated and negligible residual structure can be detected.

Positron emission tomography (PET) is a widely used molecular imaging technology. However, the inability of conventional PET systems without depth of interaction (DOI) information to precisely locate gamma rays leads to parallax error, which further results in the non-uniform resolution in reconstructed images. The existing methods enable PET systems to acquire DOI information by adding more hardwares, which are generally at the cost of higher prices and degradation of other performances. To overcome these shortcomings, we proposed a novel distance-driven cascade framework containing a bi-directional long short-term memory (Bi-LSTM) module and an encoder-decoder module. Especially, the distance-driven preprocessing was realized by splitting the sinogram into one-dimensional vectors according to radial distance and inputting them sequentially. In this approach, the bins in the same sinogram row had related features,thus were processed at the same time. Furthermore, sinogram rows used the mutual implicit information extracted by Bi-LSTM to achieve better transformation before processed by the encoder-decoder module. To test the proposed method, we conducted the network training and testing on a dataset simulated using the open-source Geant4 toolkit GATE. Compared to the DeepPET, which is a typical PET reconstruction method based on deep learning, our method acquired an obvious promotion in structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) on the test dataset. It proves that our method is superior in perceptual performance and efficiency.

Low-dose computed tomography (CT) is of great potential advantage for disease diagnosis. Usually, paired training datasets are difficult to obtain in clinical routine, which catalyzes the development of unsupervised learning techniques to improve the low-dose CT imaging. Recently, most existing unsupervised learning approaches for low-dose CT imaging were developed in the image domain, and only a few approaches have been developed in the sinogram domain, which is a challenging task. In this paper, we propose a dedicated unpaired learning technique for low-dose CT sinogram restoration with a novel data-dependent noise-generative model. The general idea is to construct a paired pseudo normal-/low-dose sinogram dataset based on the existing unpaired normal-/low-dose sinogram dataset, after which a sinogram restoration network can be obtained by training on the paired pseudo normal-/low-dose sinogram dataset. However, the difficulty of the presented idea lies in the construction of the pseudo low-dose sinogram generative network, due to the complexity of the texture feature and noise property in the sinogram domain. To address this issue, we construct an appropriative generative network architecture based on a reasonable noise-generative model in the sinogram domain, which can be used to produce pseudo low-dose sinogram data within an adversarial learning framework. To validate the proposed technique, a clinical dataset was adopted. Experimental results demonstrate that the proposed method can produce promising pseudo low-dose sinogram data, which is sufficient to train an effective sinogram restoration network. Both quantitative and qualitative measurements show that the proposed method can obtained promising low-dose CT imaging performance.

Cardiovascular disease diagnosis relies heavily on diagnostic imaging. Advancement in computed tomography (CT) technology has particularly improved diagnosis in patients with coronary artery disease. In particular, the improved spatial resolution and iodine quantification capabilities of photon-counting CT (PCCT) have the potential to further improve the diagnostic workflow. Since iodine quantification has become a critical aspect of clinical diagnosis, several studies have been conducted to evaluate its effectiveness and the parameters that may affect it. An additional relationship, the effect of spatial resolution and vessel size on iodine quantification, was examined with a designed phantom. A phantom consisted of six different tubes of changing diameters (2 to 12 mm), along with a cone and an hourglass-shaped tube with diameters from 3 to 8 mm. It was scanned on a PCCT after being filled with an iodine solution. Iodine density maps, VNC, and VMI 70keV were then reconstructed with different fields of view (250 mm, 350 mm, 450 mm). Regions of interest were placed on spectral results along the length of the hourglass. Spectral results were highly accurate for vessels larger than 4 mm in diameter and regions of interest larger than 3 mm. The bias in iodine quantification increases with smaller diameters. Conversely, VNC increased, illustrating a directly proportional relationship between VNC and iodine density. The proposed phantom design allows for future studies that further investigate the relationship between spatial resolution and iodine quantification, especially in clinical workflow for optimizing protocols, implementing new CT technologies, and harmonizing protocols between different CT platforms.

Dual-source photon-counting computed tomography (PCCT) enables a novel ultra-high-resolution (UHR) scanning mode that can provide UHR conventional images (0.2 mm) as well as spectral results (0.4 mm). To evaluate the spatial resolution and quantitative capabilities of the UHR mode, with a focus on thoracic imaging, a PixelPrint lung phantom mimicking interstitial lung disease with honeycombing and iodine rods of different diameters and concentrations directly attached to the phantom were scanned at doses from 1.0 to 7.5 mGy. Virtual monoenergetic images at 50 keV, virtual non-contrast, and iodine density maps at 0.4 and 1 mm slice thickness were reconstructed as well as conventional images at 0.2, 0.4, and 1 mm slice thickness, all with standard lung and quantitative reconstruction kernels (matrix size 512x512). Iodine quantification was performed for the attached rods, and clinically relevant features in the lung phantom were utilized to evaluate spatial resolution. Overall, iodine quantification was stable across radiation dose, reconstruction kernels, and slice thickness with errors of 0.25, 0.20, and 0.40 mg/mL for 1, 2, and 5 mg/mL iodine, respectively. Even the smallest iodine core rod was detected in the extended CT phantom for a higher dose. For the diseased lung region, images at 0.2 mm slice thickness appeared sharper and depicted smaller structures better, even with increased noise in comparison to thicker slices. In conclusion, UHR mode demonstrated high spatial resolution with detection of small features and accurate iodine quantification, which may provide diagnostic advantage to thoracic imaging with more precise and accurate information.

This study investigates the effect of radiation dose reduction of a renal perfusion CT protocol on quantitative imaging features for patients of different sizes. Our findings indicate that the impact of dose reduction is significantly different between patients of different sizes for standard deviation, entropy, and GLCM joint average at all dose levels evaluated, and for mean at the lowest dose level evaluated (p < .001). These results suggest that a size-based scanning protocol may be needed to provide quantitative results that are robust with respect to patient size.

Purpose: Legislation imposes the verification of compliance with acceptability criteria of (new) x-ray devices at their first installation in a hospital, and from then annually. In this study, QC measurements obtained from an emerging photon counting CT scanner (PCCT) are compared to 53 state-of-the-art CT scanners in routine clinical use.

Material and methods: The Belgian acceptance test protocol is applied by a team of three medical physicists on 53 scanners (33 distinctive models, 5 manufacturers) as well as on the PCCT (Naeotom, Siemens-Healthineers, Germany). Intersystem comparisons were achieved by pooling the QC data of 213 annual tests between 2016 and 2021 into a central relational database (DB) that was newly developed to improve the quality of reports and interpretation. Several test results will be illustrated in present study, including the spatial resolution from full width half maximum of a 0.28 mm bead, the accuracy of Hounsfield Unit of water using filtered back projection and iterative reconstruction, and longitudinal follow-up of CTDI_vol for 120 kVp/100 mAs.

Results: The database is a flexible, user-friendly platform that shows typical results for CT scanners of different make and model. It confirms the stability of most CT scanners and the choice of limiting values in current CT protocols. The new PCCT compared favorably to the other scanners and passed all acceptability criteria.

Conclusion: The confirmed quality of data acquisition and processing provides new insights into the settings and stability of routine CT and PCCT scanners. Protocols should now aim for an absolute performance score.

Angioplasty is an interventional treatment for blood vessel stenosis, that uses a catheter navigated to obstructed vasculature under fluoroscopy to force open the blockage with a balloon or permanent wire stent. Accurate stent placement is critical to avoid stent shifting and enabling new blockages. Unfortunately, background structures such as bone and tissue interfere with stent visualization during angioplasty, making it difficult for physicians to visualize stents during placement and post-operative checkups. This contributes to the 25% restenosis rate in patients with permanent stents¹. Multi-energy subtraction angiography (ESA) with fast kV switching may enhance nickel-stent visualization by suppressing background structures. The objective of this work was to develop an ESA protocol to enhance the visualization of nickel wire stents. The signal-to-noise ratio of nickel for a given radiation exposure was simulated using MATLAB® and validated against images of a deployed nitinol stent overlaid on a chest phantom. The stent was isolated from bone or soft tissue by applying a calculated weighting factor to the low energy image. We found that by manipulating the weighting factor, it is possible to selectively remove anatomical structures obstructing the view of the stent.

Magnetic resonance imaging (MRI), a non-invasive and safe imaging method, is largely used for the assessment of multiple organs including knee. Due to the complexity of the knee joint, better images quality are required. Over the last decades, a variety of coil solutions has been proposed to improve MR image quality. Interestingly, dielectric or metamaterial structures have been used as additional devices for their ability to tailor electromagnetic field at a given scale. However, the use of these devices is often limited by their complexity and bulkiness. The present study aimed at improving the B1 transmit field for knee imaging at 3T through the design and manufacturing of a convenient and comfortable passive metasurface. A cylindrical array of conductive stripes was used to redistribute by inductive coupling the radiofrequency field generated by the body coil of the MRI scanner. This design takes no more space than a thin sheet placed around the leg of the patient. We have shown in simulation and experimentally the accuracy of this solution. For a given flip angle during signal acquisition, the improved transmit field allowed a reduction of the necessary input power. In addition to that, the structure had a negligible influence on the electric field inside the tissue and so did not significantly increase the specific absorption rate (SAR).

In most clinical image-guided procedures like stenting and needle based biopsies one is only interested to visualize a specific part of the patient’s body. Therefore, only the relevant parts should be irradiated in order to minimise the patient’s dosage. Utilizing such volume-of-interest imaging has the potential to save a lot of dose but causes artifacts to appear in the resulting image, stemming from the reconstruction from incomplete projections. Modern bi-plane C-arm CT systems have the potential to acquire some projections complete and other projections limited to the region of interest. Here, we investigate how extrapolation methods can be used to replace the missing data and minimize the resulting artifacts. A new extrapolation method is proposed and compared to an existing extrapolation method. Subsequently, both the image quality in the recorded region of interest as well as the extrapolated outer area are examined. It is shown, that blockwise contiguous truncated/non-truncated bi-plane projections are an advantageous compromise between fully acquired data and completely truncated data. Significant amounts of dosage can be saved, whilst largely maintaining the relevant image quality.

Alveolar wall extraction is important to elucidate the 3D microstructure of alveolar ducts and alveolar sacs. The purpose of this study is to improve the accuracy and speed of alveolar wall extraction using U-Net, and to contribute to the analysis of the 3D lung microstructure. In our experiments, we first performed alveolar wall extraction using 2D U-Net. As a result, the accuracy rate and F-measure were 0.982 and 0.916, respectively. Next, the model created by 2D U-Net was tested on the axial, coronal, and sagittal planes of the 3D image, and the results were used to create training data for 3D U-Net by performing AND, Majority-Vote, and OR operations. The results of the Majority-Vote operation yielded an accuracy rate of 0.980 and an F-measure of 0.909, which was the best among the three operations. In addition to the metrics, we also evaluated the accuracy in terms of topology. The results of the 2D U-Net test were 12 connected components, 17 cavities, and 164 holes, while the results of the Majority-Vote operation were counted as 36 connected components, 35 cavities, and 105 holes. Since these numbers alone are insufficient to determine which is superior, a qualitative evaluation is also necessary. Qualitative evaluation showed that some images were more accurate as a result of Majority-Vote operation than as a result of 2D U-Net.

Self-supervised learning for CT image denoising is a promising technique because it does not require clean target data that are usually unavailable in the clinic. Noise2void (N2V) is one of the famous methods to denoise the image without paired target data and it has been used to denoise optical images and also medical images such as MRI, and CT. However, the performance of the N2V is still limited due to the restricted receptive field of the network and it decreases the prediction performance for CT images that have complex image context and non-uniform Poisson random noise. Thus, we proposed enhanced N2V that utilizes penalty-driven network optimization to further denoise the images while preserving the important details. We used the total variation term to further denoise the image and also the laplacian pyramids term to preserve the important edges of the image. The degree of the influence of each penalty term is controlled by the hyperparameter value and they are optimized to achieve the best image quality in terms of noise level and structure sharpness. For the experiment, the real dental CBCT projection data were used to train the network in the projection domain. After the network training, the test results were reconstructed and compared at each different dose level. Meanwhile, PSNR, SNR, and a line profile were also evaluated to quantitatively compare the original FDK images, and proposed method. In conclusion, the proposed method achieved further denoises the image than N2V even preserving the details. By penalty-driven optimization, the network was able to learn the spectral features of the image while still the receptive field is limited to avoid identity mapping. We hope that our method would increase the practical utility of network-based CT images denoising that usually the target data are unavailable.

A Monte-Carlo model was developed to simulate the response of a pixelated hyperspectral CZT X-ray detector. The first part of the simulation was carried out using Geant4, to obtain a list of energy depositions inside the CZT crystal. The second part of the simulation used charge transport equations to calculate the size of the electron charge cloud, as it drifts under the electric field to be read out. Experimentally acquired data from an Am-241 source with the HEXITEC detector were compared to simulated data, and good agreement was found. The model was used to investigate the energy dependence of fluorescence and charge sharing effects. Firstly, the probability of producing an escaped fluorescence photon was quantified as a function of primary photon energy. As expected, at primary photon energies just above the K-edge of Cd, there is a greater chance of producing an escaped fluorescence photon, and this probability decreases as the primary photon energy increases. Secondly, the probability of an event being shared across multiple pixels as a function of primary photon energy was quantified. It was found that as the primary photon energy is increased, there is a greater chance of producing an event shared across multiple pixels. The detector response to a Bremsstrahlung spectrum was simulated. Using previous results, fluorescence and charge sharing effects were corrected for, giving a corrected spectrum in good agreement with the input spectrum.

Dual-energy (DE) thoracic imaging with photon counting detectors (PCDs) enables suppressing anatomic noise due to bone structures, potentially improving the detection of lung nodules. Relative to DE approaches based on the kV-switching method, a PCD-based, single-exposure approach may reduce the potential for motion artifacts and suppress electronic noise. In this work, we experimentally optimize the energy thresholds and tube voltages for single-exposure DE thoracic imaging implemented with PCDs. We used a phantom consisting of uniform acrylic slabs totaling 8.6 cm in thickness with a 16 cm air gap between slabs. A lung nodule was simulated by adding 2 mm of acrylic in a small region of interest in the center of the phantom. We imaged the phantom using a PCD with a 750-μm-thick cadmium telluride x-ray convertor, 100×100 μm² detector elements, two user-defined energy thresholds, and analog charge summing for charge sharing correction. We acquire images for tube voltages ranging from 90 kV to 130 kV in 10-kV increments. For each tube voltage, we varied the energy threshold from 32 keV to 90% of the maximum beam energy. The contrast-to-noise ratio (CNR) of the simulated nodule was measured for each combination of tube voltage and energy threshold. Optimal energy thresholds range from 45 keV to 55 keV for tube voltages ranging from 110 kV to 120 kV. When optimized for the energy thresholds, there was relatively little variability in the CNR across tube voltages beyond that due to measurement error. Future work will compare the CNR of PCD-based DE radiography with the conventional kV-switching approach.

Quantitative angiography (QAngio) may provide hemodynamic information during neurointerventional procedures through imaging biomarkers related to contrast flow. The standard clinical implementation of QAngio is limited by projection imaging: analysis of contrast motion within complex 3D geometries is restricted to 1-2 projection views, truncating the potential wealth of imaging biomarkers related to disease progression or efficacy of treatment. To understand the limitations of 2D biomarkers, we propose the use of in-silico contrast distributions to investigate the potential benefits of 3D-QAngio within the context of neurovascular hemodynamics. Ground-truth in-silico contrast distributions were generated in two patient-specific intracranial aneurysm models, accounting for the physical interactions of contrast media and blood. A short bolus of contrast was utilized to obtain full a wash-in/ wash-out cycle within the aneurysm ROI. Simulated angiograms mimicking clinical cone-beam CT (CBCT) acquisitions were then generated, and volumetric contrast distributions were reconstructed to analyze bulk contrast flow. The ground-truth 3D-CFD, reconstructed 3D-CBCT-DSA, and 2D-DSA projections were used to extract QAngio parameters related to contrast time dilution curves, such as area under the curve (AUC), peak height (PH), mean-transit-time (MTT), time-to-peak (TTP), and time to arrival (TTA). An initial comparison of quantitative flow parameters in both 2D and 3D, in a smaller and larger aneurysm, indicated that 3D-QAngio can provide a good description of bulk flow characteristics (TTA, TTP, MTT), but recovery of integral parameters (PH, AUC) aneurysms is limited. Nonetheless, incorporation of 3D-QAngio methods may provide additional insight into our understanding of abnormal vascular flow patterns.

Tomosynthesis acquires projections over a limited angular range, resulting in anisotropic sampling in the Fourier domain. The volume of the sampled space is therefore spatially dependent; different Fourier components are sampled for the same object, depending upon where the object is located relative to the system origin. A next-generation tomosynthesis (NGT) system was developed at the University of Pennsylvania to increase the spatial isotropy in DBT, by incorporating additional system motions. In this work, we investigate the spatial dependency of image quality in tomosynthesis and compare conventional and NGT tomosynthesis in terms of multiplanar reconstruction (MPR). Two test objects, a high-frequency star pattern and a low-frequency octagon phantom, were placed throughout the detector field of view at various obliquities to analyze the anisotropic nature of tomosynthesis. Reconstructions of the star pattern were analyzed both qualitatively and quantitatively using the Fourier distortion metric (FSD). Reconstructions of the octagon phantom were analyzed qualitatively. In a separate experiment, a container filled with water and acrylic beads of various diameters were imaged at various locations to simulate low-contrast objects mimicking breast tissue. We show that the spatial dependency of MPR is unique to the tilt angle, orientation, and frequency of the input. The NGT geometry benefitted the visualization of objects by reducing the out-of-plane artifacts in MPR.

In regularized PET image reconstruction, the performance of a reconstruction algorithm depends crucially on the smoothing parameter that controls the balance between the likelihood term and the regularization term. In this work, we propose a new method of tuning the smoothing parameter using deep learning. Unlike the traditional estimation-theoretic approach where the smoothing parameter is estimated directly from the observed noisy data, the deep learning-based method requires large amounts of prior training pairs, which are usually unavailable in routine clinical practice. To overcome this problem, we extend our previous work on the training-set approach which provides a mathematical formula to calculate the smoothing parameter for the simple quadratic spline regularizer. We note that this training-set approach is effective only when the noiseless representative exemplars are available. For deep learning, however, collecting large amounts of such noiseless exemplars for a training dataset is unrealizable. Therefore, instead of collecting them, we generate diverse images similar to the exemplars of the ground truth radiotracer distribution using Gibbs sampling followed by post-processing and calculate the smoothing parameter value for each image. For our deep learning model, we use the residual network architecture that allows deeper layers with higher efficiency than the typical convolutional neural network. The experimental results show that our deep learning method provides optimal values of the smoothing parameter, which are comparable to the accurate values calculated from noiseless exemplars, for a wide range of underlying images.

Since the multi-tracer positron emission tomography (PET) technology will increase the total dose of tracers, it is necessary to study the low-dose multi-tracer PET technology. In this paper, we proposed a model based on attention mechanism to estimate the standard-dose single-tracer sinogram from the low-dose dual-tracer sinogram, that is, to achieve both standard-dose estimation and dual-tracer signal separation in sinogram domain. Both the spatial attention and channel attention were implemented. We cascaded the FBP-Net for reconstruction after the proposed model, and verified the proposed method in simulation experiment and rat experiment. At the same time, we compared the results of dual-tracer PET signal separation in sinogram domain and activity map domain. The results prove the effectiveness of the proposed method in the problem of low-dose dual-tracer PET image reconstruction, and also show that signal separation in sinogram domain is more effective than separation in activity map domain.

In medical imaging, material decomposition and lesion differentiation are essential for the early detection of many severe and deadly diseases. While standard flat panel imagers are capable of providing high-resolution images, they lack the ability to differentiate between X-ray energies and, therefore, soft and hard tissues. Detection of multiple X-ray energies by dual-shot, photon counting, or dual-layer detectors provides a route for differentiation of these tissues based on the spectrum observed; multiple images may be generated based on the energies, allowing for subtraction and a more detailed image of tissues and calcifications. To improve the resolution, dose level, and motion artifacts of these imagers, we propose a new dual-layer detector consisting of a direct conversion amorphous selenium top layer, followed by an indirect conversion scintillator/amorphous selenium bottom layer. In this work, we present the first steps towards building this detector by characterizing the performance of the bottom indirect flat panel. We show that the blocking layer chosen, though unoptimized, performs adequately up to 50 V/um and have fabricated the detector, which will be evaluated for detective quantum efficiency and modulation transfer function.

Head magnetic resonance imaging (MRI) is susceptible to motion artifacts. Images with gross motion artifacts cannot be easily corrected, which makes the monitoring of patient movement is essential to improve the acquired image quality. Several methods have been proposed to detect patient movement during MRI scanning, such as navigator, optical tracking system and the image quality assessment. However, they have some disadvantages such as insufficient real-time performance and requiring additional MR data acquisition. Herein, we propose an MR-compatible millimeter wave (mmWave) radar system to monitor some typical types of movements during head MRI scanning. The radar sensor is installed in a 3T MR system. When MRI system scanning, the radar installed will vibrate. A novel algorithm is proposed to generate the spectrogram without vibration interference. Due to limited movement, we explore how the range information rather than micro-Doppler information can assist the activity classification. After data processing, convolutional neural network (CNN) is applied to extract the features in multiple range bins. The extracted features will be inputted to a shallow neural network for recognizing the moving parts. The result shows four different activities are classified with an overall accuracy of (96.81±0.69)%. Finally, experimental results with acquired head MR images prove our system’s sufficient capability of recognizing nodding head yet inadequate capability of detecting head shaking movement.

This paper implements a generative adversarial deep learning network (GAN) to automate the generation of realistic and representative 3D cancer models to investigate digital breast tomosynthesis (DBT) with virtual clinical trials (VCT). Initially, a series of mass lesions from wide-angle DBT cancer cases were manually segmented. We trained a 3D-GAN in two phases: the first phase utilized 105 manually created models of invasive ductal carcinoma (IDA) (including both microlobulated and spiculated lesions) as a training dataset; the second phase focused on enhancing the details of the borders of the GAN-models by using a smaller training set of 42 highly spiculated segmentations. To improve the realism of the generated models in VCTs from both phases, post-processing was carried out by removing disconnected pixels, filling holes, smoothing the borders and elongating existing spicules. Fifteen generated lesions were then simulated in acquired patient images and their realism was validated by a radiologist. 80% of the simulated cases received at least a realism score of 3 out of 5. While the average realism score of the generated voxel models was slightly lower than the average score of manually segmented lesions, the 3D-GAN successfully generated breast cancer models of spiculated masses even when trained with a limited dataset. This same method could be applied to generate other mass models of certain subgroups to allow lesion specific simulations, increasing the efficiency of the process to produce representative lesion models for VCTs.

Xenon-enhanced, dual-energy x-ray radiography has been proposed for imaging of lung ventilation. It is important to assess the ability of dual-energy subtraction to suppress anatomic noise associated with lung parenchyma. Anatomic noise in thoracic radiography obeys an inverse power law and there exist imaging phantoms that mimic this power law. Such phantoms are based on a random, tight packing of solid acrylic spheres and are not suitable for lung ventilation studies. We developed a phantom based on randomly-packed, hollow acrylic cylinders with inner diameters of 1.59 cm, wall thicknesses of 0.16 cm and lengths of 1.59, 1.27, 0.95, 0.64 or 0.32 cm. The number of segments of each length was chosen to approximately match the volume of space occupied by each set of segments. Measurements of the effective density of the packed cylinders yielded ~0.26 g cm^-3. A randomly-packed-sphere phantom was also constructed as a reference. Both phantoms were imaged using a flat-panel detector at tube voltages of 50 kV to 150 kV. A power-law model (NPS ∝ κ/|u|^β) was fit to the anatomic noise power spectra. The β-value of the cylinder phantom was within 1/5 of that of the sphere phantom, although both phantoms yielded power-law parameters ranging from 2.0 to 2.4, which is lower than that reported in the literature. The κ-value of the cylinder phantom was ~1.1 times that of the sphere phantom. We conclude that the cylinder-based clutter phantom, with some modifications, can be used to simulate the anatomic noise power spectrum in thoracic radiography.

Approximately 2.5% of the proton range uncertainty comes from computed tomography (CT) number to material characteristic conversion. We aim to conquer this CT-to-material conversion error by proposing a multimodal imaging framework to enable deep learning (DL)-based material mass density inference using dual-energy CT (DECT) and magnetic resonance imaging (MRI). To ensure the robustness of DL models, we integrated physics insights into the framework to regularize DL models and achieve DL using small datasets. Five MRI-compatible phantoms were created from tissue-mimicking materials that served as a ground true reference to validate the proposed framework. The reference mass densities for each phantom were measured by a 150 MeV proton beam. Multimodal images were acquired from T1- and T2-weighted images and DECT images as training and validation data for DL. Residual networks (ResNet) were implemented to evaluate the feasibility of the proposed framework. ResNet-DE-MR denotes that ResNet was trained with MRI and DECT images, while ResNet-DE presents that only DECT images were used to train ResNet. ResNet was also compared to an empirical DECT model. Meanwhile, a retrospective patient case was included in the study to demonstrate the proof of concept for the proposed framework. The phantom validation experiment showed that ResNet-DE-MR achieved mass density errors of -0.4%, 0.3%, 0.4%, 0.7%, and -0.2% for adipose, muscle, liver, skin, and bone. The proposed DL-based multimodal imaging framework was demonstrated to enable accurate material mass density inference using DECT and MR images. The framework can potentially improve the treatment quality for proton therapy by reducing proton range uncertainty.

The current clinical practice for Monte Carlo (MC) treatment planning reserves a 3.5% margin to compensate for proton range uncertainty. Additionally, patient positional uncertainty is typically 3-5 mm for proton craniospinal irradiation (CSI) treatment planning. These two uncertainties compromise the sparing of spine vertebrae in proton CSI patients. Computer tomography (CT) material characterization contributes approximately 2.5% proton range uncertainty. Multiple CT-tomaterial conversion methods have been investigated using dual-energy CT or magnetic resonance imaging to improve the range uncertainty. However, there is a lack of experimental data to validate the credibility of those material characterization models. We aim to develop an in vivo proton range method using pseudo proton radiography to validate imaging-based material characterization models consistently. Proton radiography techniques, such as proton water equivalent thickness (WET) and dose maps, were used to evaluate the in vivo proton range accuracy. Anteroposterior proton beams were penetrated through an anthropomorphic phantom. Then the exit doses were measured from proton radiography imaging. The validation experiment applied a newly designed multi-layer strip ionization chamber (MLSIC) for the first time to perform four-dimensional (4D) measurement for depth doses from 625 proton spots in two minutes. The depth doses of each spot were post-processed into WET imaging. A MatriXX PT was applied for 2D measurement from 19x19 cm2 proton fields. We compared the performance of the empirical DECT model and physics-informed machine learning (PIML) models for material conversion. The results indicated that the PIML-based material characteristic method generated more accurate WET and dose imaging using DECT compared to conventional machine learning and empirical material inference methods. The proposed in vivo proton range validation method can be used to quantify the credibility of DECT-based material conversion models for proton range enhancement. The method can potentially provide in-room patient anatomy changes to accomplish online adaption for modification. This technique will significantly benefit proton flash therapy, which demands high accuracy.

Computed tomography (CT) imaging is a key element of effective radiation therapy treatment, providing a threedimensional patient model for treatment planning dose calculations and precise patient positioning at the time of treatment. For radiotherapy, CT images are acquired using either an additional kilovoltage (kV) x-ray source mounted onto the treatment system or the system’s intrinsic megavoltage (MV) source. With beam geometry similar to fan-beam CT, tomographic therapy systems have unique potential for performing MV-kV dual-energy CT (DECT) imaging. This work was undertaken to quantify the prospects of MV-kV DECT. Six MV-kV and three kV-kV spectral combinations were considered. Single ray signal-to-noise ratios (SNRs) were calculated for a two-material object using estimation theory in the context of basis-material decomposition. Maximum-achievable SNR and optimal dose allocation were compared across spectral combinations and object thicknesses. Basis material images of the XCAT phantom were simulated using 140kV-80kV and DetunedMV-80kV spectral pairs, using both optimized and equal dose distributions. Results demonstrated that optimal SNR is achieved with >80% dose allocated to the MV spectrum (1-cm bone, 40-cm tissue object). SNR improved by as much as 31.5% by optimizing the dose allocation. This yielded 26% noise reduction and 23% SNR improvement in the XCAT basis material images. Peak SNR and optimal dose distribution varied considerably for different object thicknesses, suggesting MV-kV DECT protocol should be carefully chosen for a given imaging task. This study demonstrates the potential for enhanced contrast with MV-kV DECT imaging and warrants further investigation into its practical implementation in radiotherapy settings.

Improving the modeling of the breast shapes during mechanical compression in both cranio-caudal (CC) and medio-lateral oblique (MLO) views can enhance the development of image processing and dosimetric estimates in digital mammography and digital breast tomosynthesis (DBT). In previous work, a CC model was created using a pair of optical structured light scanning systems, but acquiring similar data for an MLO view model during clinical practice proved impractical with these devices. The present work instead uses two smartphone infrared cameras with 3D-printed holders to obtain surface scans for the MLO view during a DBT acquisition. The study compared the average distance between the MLO breast shape information recorded by the smartphone-based scans to the corresponding DBT exam-based surface scan for 20 patient breasts. Results showed that there was close overlap between the smartphone-based scanned surfaces of the breast and the corresponding DBT images. The agreement between the breast shape represented by these surfaces was dependent on the smartphone-based scanner precision, the segmentation procedure used to obtain the DBT surface, and the manual alignment of the smartphone-based left and right-side view of the breast. The agreement was however of sufficiently good quality for the data to be used for the development of an MLO breast shape model.

Digital Breast Tomosynthesis (DBT) is a medical imaging modality that has been increasingly used for breast cancer screening. To improve the accuracy in the early detection of breast cancer, it is common to use tools based on image processing to improve the quality of DBT images and, consequently, the visibility of lesions of clinical interest. Microcalcification (MC) clusters are in the class of findings that may indicate the early stages of breast cancer. To evaluate the impact of image processing methods in the detection of breast lesions, human perception studies are usually performed, where detection accuracy is evaluated using positive and negative cases, i.e., images with and without lesions. However, only a small fraction of clinical cases have positive cases and the exact location of the lesion is not always known. Thus, in this work, we present a method to artificially insert MC cluster in normal exams of DBT. A set of MCs was segmented from clinical cases acquired in a prone stereotactic breast biopsy system. In the pipeline, we built a MC cluster and insert it in a 3D volume. Then, we project it on the detector and inserted the MC cluster into the projections of a clinical DBT exam. It is possible to define the size, contrast, and location of the MC cluster accordingly. We performed a two-alternative forced-choice (2-AFC) study with six experienced medical physicists and the average success rate was 53.83%, suggesting that readers, on average, could not distinguish between real and simulated MCs. Overall results showed that images with simulated MC are similar to the real clinical cases. Our source code is available at www.github.com/LAVI-USP/MCInsertionPackage-DBT.

Fluoroscopy in a low-dose tube output is used to reduce the damage associated with radiation exposure. However, lowering the radiation dose inevitably increases random noise in x-ray images, resulting in poor diagnostic image quality, which requires noise reduction for accurate diagnosis. Also, in the case of non-static objects, the image is blurred due to motion. The most-used denoiser with a recursive filter (RF) preserves details well when applied to temporal data, but it is vulnerable to motion blur. Existing convolutional neural network (CNN)-based algorithms with single-frame input cannot use the temporary context, and others with multi-frame input are good for motion detection but poor for detail preservation. Therefore, we propose a motion-level-aware denoising framework to combine the results of RF- and CNN-based algorithms depending on the pixel-wise magnitude of motion to complement each other. The data we use are fluoroscopy images taken in continuous time, and we aim at many-to-one so that one frame is denoised by considering sequential frames. Also, since both RF- and CNN-based algorithms used in our architecture are many-to-one methods, they can consider spatiotemporal information. In the multi-frame input, the difference in intensity of each pixel between frames is calculated to obtain a moving map. Depending on the factor value from the moving map, the final image is obtained by reflecting the outputs of the RF- and CNN-based algorithms. If the factor value is high, the pixel intensity of the final image is like the CNN-based output, which is good for motion detection, and vice versa, it more reflects the intensity of RF output, which is excellent in perceptual quality. Therefore, it prevents motion blur and does not over-smooth microdetails, such as bones and muscles. The results show that combining the two outputs together records higher peak signal-to-noise ratio (PSNR) and has better perceptual quality for diagnosis than using only one method. Furthermore, our combining method can output x-ray images of higher quality by using more advanced networks in future fluoroscopy denoising studies, since the proposed denoising framework is not only applicable to specific architectures used in this study but can also be broadly applied to other alternative networks.

So far, dual-tracer positron emission tomography (PET) imaging has been a rising topic in the field of PET imaging. This study focused on the reconstruction and signal separation for simultaneous triple-tracer PET imaging, where three tracers were injected at the same time. To reconstruct the image of each tracer, we proposed a three-dimensional encoder-decoder network based on multi-task learning. The mixed dynamic sinogram of three tracers was input into the encoder module. Then, the three different decoding modules output the dynamic image of each tracer according to their unique characteristics. The proposed model could simultaneously learn the spatial information and temporal information from the mixed PET signals. The reconstructions were evaluated by multi-scale structural similarity (MS-SSIM) and peak signal-to-noise ratio (PSNR). The robustness of this method was verified by simulated datasets with different phantoms, tracer combinations and sampling protocols.

An important feature enabled by Photon-Counting Detector (PCD) CT is the simultaneous acquisition of multi-energy data, which can produce virtual monoenergetic images (VMIs) at a high spatial resolution. However, noise levels observed in the high-resolution (HR) VMIs are markedly increased. Recent work involving deep learning methods has shown great potential in CT image denoising. Many CNN applications involve training using spatially co-registered low- and high-dose CT images featuring high and low image noise, respectively. However, this is implausible in routine clinical practice. Further, typical denoising methods treat each VMI energy level independently, without consideration of the valuable information in the spectral domain. To overcome these obstacles, we propose a prior knowledge-aware iterative denoising neural network (PKAID-Net). PKAID-Net offers two major benefits: first, it utilizes spectral information by including a lower-noise VMI as a prior input; and second, it iteratively constructs refined datasets for neural network training to improve the denoising performance. This study includes 10 patient coronary CT angiography (CTA) exams acquired on a clinical HR PCD-CT (NAEOTOM Alpha, Siemens Healthineers). The HR VMIs were reconstructed at 50 and 70 keV, using a sharp kernel (Bv68) and thin (0.6 mm, 0.3 mm increment) slice thickness. Results showed that the PKAID-Net provided a noise reduction of 96% and 70% relative to FBP and iterative reconstruction, respectively while maintaining spatial and spectral fidelity and a natural noise texture. These results demonstrate the noise reduction capacity of PKAID-Net as applied to cutting-edge PCD-CT data to enable HR, multi-energy cardiac CT imaging.

Flat-panel, cadmium telluride (CdTe) photon counting detectors (PCDs) are of interest for contrast-enhanced spectral mammography (CESM). Modelling of the frequency-dependent imaging performance of CdTe PCDs is useful for optimization of technique parameters. The purpose of this work is to experimentally validate a Monte-Carlo-based model of the frequency-dependent imaging performance of a CdTe PCD for contrast-enhanced breast imaging. Our Monte-Carlo model accounts for the basic physics of x-ray detection by PCDs, including the finite range of photo-electrons, characteristic emission and subsequent reabsorption, the finite size of charge clouds due to diffusion and Coulomb repulsion, electronic noise and energy thresholding. We validated the model using a two-bin CdTe PCD with charge summation for charge-sharing correction. The detector consists of a 750-um-thick CdTe converter with 100 um pixels. Our model was used to predict both the modulation transfer function (MTF) and the normalized noise power spectrum (NNPS) of both energy bins, in addition to the cross NPS between energy bins. Validation was performed for 40 kVp, 45 kVp and 50 kVp tube voltages. For each tube voltage, we validated for energy thresholds that resulted in photon ratios between the two energy bins of 1:2, 1:1 and 2:1. In all cases, our model predicted the MTF of both energy bins within 3%. The modelled NNPS was 5% to 15% of the measured NNPS. We conclude that the Monte-Carlo-based model can be used to model the frequency-dependent signal and noise properties of contrast-enhanced spectral mammography implemented with CdTe PCDs.

This conference presentation was prepared for the Medical Imaging 2023: Physics of Medical Imaging conference at SPIE Medical Imaging 2023.

Preliminary results have shown replacing the conventional x-ray tube with a stack of narrowly collimated beams along the axial direction reduces scatter and improves the spatial uniformity and the CT number accuracy in CBCT. The purpose of this study was to evaluate a carbon nanotube (CNT) x-ray source array designed for a prototype multisource CBCT (ms-CBCT) for maxillofacial imaging. The CNT x-ray source array comprises 8 evenly distributed focal spots (“sources”), and 8 corresponding CNT cathodes and gate electrodes, with an inter-focal spot spacing of 12mm. A multisource collimator was designed to confine the radiation from each focal spot to a narrow cone beam covering a section of the object in the axial direction. The focal spot sizes (FSS) were measured using a pinhole camera setup following the IEC standard procedure. The radiation fields of the collimated beams were measured using a flat panel detector. The x-ray source cathode and tube currents were recorded, and radiation dose was measured with a dose meter. Preliminary 3D volumetric imaging of an anthropomorphic head phantom was conducted using the CNT source array. The targeted 15mA tube current was achieved from each focal spot at 110kVp. The measured FSS IEC nominal value was 0.7. The average x-ray cone angle of each collimated beam was 2.3°. The prototype CNT x-ray source array achieved the specifications for the oral and maxillofacial ms-CBCT scanner. Work is in progress to evaluate the imaging system.

Lens dose can be high during neuro-interventional procedures, increasing the risk of cataractogenesis. Although beam collimation can be effective in reducing lens dose, it also restricts the FOV. ROI imaging with a reduced-dose peripheral field permits full-field information with reduced lens dose. This work investigates the magnitude of lens-dose reduction possible with ROI imaging. EGSnrc Monte-Carlo calculations of lens dose were made for the Zubal head phantom as a function of gantry angulation and head shift from isocenter for both large and small FOV’s. The lens dose for ROI attenuators of varying transmission was simulated as the weighted sum of the lens dose from the small ROI FOV and that from the attenuated larger FOV. Image intensity and quantum mottle differences between ROI and periphery can be equalized by image processing. The lens dose varies considerably with beam angle, head shift, and field size. For both eyes, the lens-dose reduction with an ROI attenuator increases with LAO angulation, being highest for lateral projections and lowest for PA. For an attenuator with small ROI field (5 x 5 cm) and 20% transmission, the lens dose for lateral projections is reduced by about 75% compared to a full dose 10 x10 cm FOV, while the reduction ranges between 30 and 40% for PA projections. Use of ROI attenuators can substantially reduce the dose to the lens of the eye for all gantry angles and head shifts, while allowing peripheral information to be seen in a larger FOV.

Physicists generally use the Noise Power Spectra (NPS) and Standard Deviation (SD) to characterize noise properties associated with non-linear reconstruction algorithms. However, these metrics capture only first and second order statistics. The purpose of this work is to characterize the impact of the higher order statistics, commonly associated with non-linear reconstruction, on noise texture. Images of a 32cm water phantom were acquired on the Aquilion ONE Genesis Computed Tomography (CT) system and reconstructed with deep learning reconstruction (DLR), model-based iterative reconstruction (MBIR), hybrid iterative reconstruction (AIDR), and filtered backprojection (FBP). Regions of interest (ROIs) of 100x100pixels were extracted from the center of the images. Pure Gaussian noise counterpart image datasets with the same mean, SD, and NPS as each acquired data condition were also generated by convolving random white noise with the root-NPS of the acquired data. Nine naïve observers were tasked with distinguishing the acquired noise image from its pure Gaussian counterpart via a two-alternative forced choice experiment. Results showed the FBP images appeared indistinguishable from their pure Gaussian counterparts (Percent Correct=54%), while MBIR images were readily distinguishable from Gaussian ones (Percent Correct=98-100%). DLR and AIDR images were more difficult to distinguish from their pure Gaussian counterparts (Percent Correct=58-88%), than MBIR, which indicates that it is more similar in perceived texture to Gaussian noise. This work demonstrates the appearance of CT noise texture may be dependent on higher orders statistics not captured by the NPS; noise textures with identical NPS and SD can be distinguished based on non-Gaussian properties.

Proton therapy requires highly accurate dose calculation for treatment planning to ensure the doses delivered to the tumor precisely. The accuracy of mass density estimation dominates the uncertainty in proton dose calculation. This work proposed a fully connected neural network (FCNN) based framework to estimate mass density from single-energy compute tomography. The FCNN was design as 9 hidden layers and 150 hidden units and nonlinear activation function. A CIRS 062M electron density phantom was used to train FCNN, and CIRS M701 and M702 was used to evaluate the performance of models. For M701, FCNN has mean absolute percentage errors of mass density at 0.39%,0.92%,0.68%,1.57,0.92% over brain, spinal cord, soft tissue, lung, and bone. For M702, the mean absolute percentage errors of mass density estimation by FCNN are 0.89%,1.09%,0.70%,1.52% and 3.19%, respectively.

Artificial intelligence (AI)-based reconstruction is a promising method for MRI reconstruction. However, deep neural networks may exhibit instabilities in conditions that are difficult to identify with patient images. The purpose of this work is to investigate whether digital phantoms can help evaluate the performance of AI-based MRI image reconstruction. We chose AUTOMAP as an example of AI-based reconstruction method, with the network being trained with 50,000 paired patches of T1W healthy brain images and corresponding noisy k-space data. We tested the network with noisy k-space brain images, digital phantom images, and hybrid images (i.e., brain test images that contained an inserted lesion-like object). The set of brain test images was used to evaluate the global reconstruction accuracy in terms of mean squared error (MSE). The digital phantoms were designed to test image homogeneity and resolution. The hybrid images were constructed to mimic unhealthy patient for testing whether the AI reconstruction model trained with all healthy brain images would yield equal performance on abnormal brain test images. We also selected two test cases (one brain and one phantom) to quantitatively compare AI-based reconstruction and IFFT in terms of local reconstruction accuracy, which was measured by mean intensity and homogeneity of a region of interest (ROI) in a range of noise levels.

It was observed that AUTOMAP reduced noise variance on brain images within our pre-trained noise range compared to the IFFT reconstruction, but increased variance on phantoms creating inhomogeneous appearance in reconstructed phantom images. In hybrid images, similar degradation of performance was shown in the lesion-like area. Our preliminary results demonstrated that performance of the neural network was highly dependent on the training dataset. If the training data only includes healthy subjects, reconstruction of pathology regions may not be as good as healthy anatomic regions. Digital phantoms helped identify this potential generalizability issue in this AI-based MRI reconstruction.

In this study, multi-view synthetic radiographs of patients generated from stationary intraoral tomosynthesis (sIOT) images were compared with standard bitewing (SBW) radiographs in terms of proximal tooth overlap and image quality. Patient sIOT images from a previous study were used to create seven synthetic radiographs with different viewing angles through specialized code. The proximal tooth overlap, contrast, contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR) were calculated for representative patient data sets. The synthetic radiographs had a minimum overlap of 0%, median overlap of 0%, maximum overlap of 14%, and interquartile range of 0%. The SBW radiographs had a minimum overlap of 13%, median overlap of 23.5%, maximum overlap of 49%, and interquartile range of 14.5%. The ratio of mean contrast, mean CNR, and mean SNR between the synthetic radiographs and the SBW radiographs were 3.81, 2.63, and 0.821 respectively. The synthetic radiographs had decreased proximal overlap and increased contrast and CNR compared to SBW radiographs. These results suggest that synthetic radiographs can reduce proximal overlap and improve contrast compared to SBW radiographs.

Dual-energy CT (DECT) with limited-angular-range (LAR) data has the potential to reduce radiation dose, scanning time, and motion effect, and avoid collision between the moving gantry and the patient. There are two primary sources of image artifacts, LAR and beam-hardening (BH) effects. Previous works have demonstrated that LAR artifacts can be effectively reduced or eliminated by the directional-total-variation (DTV) constraints on the orthogonal axes of the image array and BH artifacts can be corrected for by using the data-domain decomposition approach with overlapping scanning arcs. In this work, we investigate a one-step method for the simultaneous correction of LAR and BH artifacts for DECT with LAR data, thus enabling flexible scanning configurations, such as completely non-overlapping scanning arcs. Specifically, two scanning configurations, two-orthogonal-arc (TOA) and two-parallel-arc (TPA) configurations, are used to generate data from a digital chest phantom with low and high-kVp spectra. Basis images are reconstructed directly from low- and high-kVp data by solving a non-convex optimization problem with DTV constraints. They can then be combined into monochromatic images for visual inspection and quantitative analysis. The results suggest that accurate monochromatic images can be obtained from TOA and TPA configurations of 90◦ arcs, and that the TOA configuration appears to be more robust to data inconsistencies such as noise.

In radiation treatment planning (RTP), CT reconstruction that combats projection truncation artifacts induced by the patient being positioned partially outside the scan field-of-view (FOV) needs to maintain high geometric and Hounsfield Unit (HU) accuracy outside the scan FOV. A new image reconstruction method has been proposed for clinical helical CT simulation scans. This method generates support images using the Discrete Algebraic Reconstruction Technique to accurately estimate patient contours outside the scan FOV and then uses support images to guide the projection extension. The proposed method improved geometric accuracy in objects outside the scan FOV compared to a more conventional method and kept the geometric distortion within 5 mm under very severe truncation. It also demonstrated HU accuracy in objects outside the scan FOV within 2.5% for a variety of soft tissues and 15% for bone tissues on a typical electron density phantom. Images of three radiotherapy patient cases reconstructed with the proposed method exhibited clearly defined, naturally looking patient contours, including the recovery of skinfold outside the scan FOV. The proposed method shows the potential for providing clinically desirable extended FOV images for a variety of patient setups in RTP.

The recorded radio-frequency signals in photoacoustic tomography (PAT) are greatly affected by the attenuation and dispersion of the acoustic waves (while propagating through the coupling medium). The frequency-dependent attenuation causes a significant reduction of high-frequency signals causing a blurring in the reconstructed images. Here, we present a simulation study involving a model-based approach that can reduce blurring significantly for such a propagating medium. The model matrix was constructed by utilizing Green’s function for acoustically lossy and dispersive medium. As far as we know, this approach has never been employed for the construction of the model matrix. A numerical phantom (containing five discs with different radii and optical absorption properties) embedded in an acoustically lossy and dispersive medium was considered in this study. The PA signals were captured by 80 detectors (uniformly placed around the sample over 0-2π). The image reconstruction was accomplished via the Tikhonov regularization (TH) method. The back-projection (BP) and interpolated time reversal (ITR) algorithms were also implemented for image creation. The numerical values of some standard image quality metrics were calculated. The BP technique fails to recover the initial pressure map of the phantom. The ITR images appear slightly blurry but it restores the nominal pressure distribution better than the BP method. The model-based method compensates for the effect of medium-dependent attenuation during the matrix-inversion operation. Its performance is comparable to the ITR method for large structures but it works better for small structures. Therefore, it may be useful for PAT imaging of blood vessels in the tissue.

Dual-energy computed tomography (DECT) is a promising imaging modality. It has the potential to quantify different material densities and plays an important role in many clinical applications. To enable multiple material decomposition (MMD), the conventional analytical MMD algorithm assumes the presence of at most three materials in each image pixel, and each pixel is decomposed into a certain basis material triplet. However, the MMD algorithm requires strong prior knowledge of the mixture composition, and the decomposition performance is compromised around the boundaries of different compositions. In this work, we developed an analytical model based deep neural network MMD-Net to achieve multi-material decomposition in DECT. In particular, the type of the basis material triplet in each image pixel and the attenuation coefficients of each material are learned by dedicated convolution neural network modules, and the material-specific density maps are obtained from the analytical MMD algorithm. Physical experiments of a pig leg and a pork backbone specimen with inserted iodine concentrations were acquired to evaluate the performance of the MMD-Net. Results show that the proposed MMD-Net could provide high decomposition accuracy, and reduce the decomposition artifacts.

Chest X-ray is one of the most commonly performed radiologic procedures for respiratory diseases. Digital tomosynthesis (DTS) provides volumetric anatomic information at lower cost and dose compared to computed tomography (CT). However, current DTS system provides insufficient patient positioning feedback and requires a large number of reconstructed slices in order to ensure imaging the entirety of targeted anatomy. We propose an anatomy registration prototype using measurements from RGB-D cameras to 1) assist acquisition workflow, 2) provide individual-specific anatomical information to improve tomosynthesis reconstruction. Our experiments show that anatomy registration can provide real-time feedback of patient 2D position and body thickness. Our reconstruction simulations show that the anatomy and body information can speed up DTS reconstruction, reduce the number of redundant tomosynthesis slices, and help reduce image interpretation time.

The purpose of this pilot study was to evaluate if a deep learning network can be used for brain segmentation of grey and white matter using spectral computed tomography (CT) images. Spectral CT has the advantage of a lower noise level and an increased soft tissue contrast, compared to conventional CT, which should make it better suited for segmentation tasks. Being able to do volumetric assessments on CT, not only magnetic resonance imaging (MRI) would be of great clinical benefit. The training set consisted of two patients and the validation data set of one patient. Included patients had a brain CT from a spectral CT as well as a T1-weighted MRI. MRI was used for an MR-based segmentation using FreeSurfer. A convolutional neural network was trained to identify grey and white matter in virtual monoenergetic images (70 keV) from spectral CT, using the MR-based segmentation as reference, and tested to assess its’ performance. The network was able to identify both grey and white matter in roughly the correct areas. In general, there was an overestimation of grey matter. These results motivate further studies, as we predict that the network will be more accurate when trained on a larger data set.

For the U-Net based low dose CT (LDCT) imaging, there remains an interesting question: can the LDCT imaging neural network trained at one image resolution be transferred and applied directly onto another LDCT imaging application of different image resolution, provided that both the noise level and the structural content are similar? To answer this question, numerical simulations are performed with high-resolution (HR) and low-resolution (LR) LDCT images having comparable noise levels. Results demonstrated that the U-Net trained with LR CT images can be used to effectively reduce the noise on HR CT images, and vice versa. However, additional artifacts may be generated when transferring the same U-Net to a different LDCT imaging task with varied image spatial resolution due to the noise induced 2D features. For example, noticeable bright spots were generated at the edges of the FOV when the HR CT image is denoised by the LR CT image trained U-Net. In conclusion, this study suggests that it is necessary to retrain the U-Net for a dedicated LDCT imaging application.

Computed Tomography (CT) is one of the most used imaging modality in clinics. However, the associated high radiation is a major concern, and then the tradeoff between clinically accepted CT images and radiation dose is desired. In recent years, inspired by deep learning techniques, various deep learning-based methods have been developed for low-dose CT imaging, and they have potential to improve image quality of low dose CT. Meanwhile, most of these methods are trained on the CT datasets with the specific imaging geometry, and they could fail to successfully reconstruct the CT images from the other imaging geometries simultaneously, which might lead to poor generalization ability. In this work, to address this issue, we propose a dual-domain modulation for high-performance multi-geometry low-dose CT image reconstruction (DM-MG) method. Specific, the proposed DM-MG consists of two sub-networks, i.e., multilayer perceptron controlling the multiple imaging geometries, and inverse Radon transform approximation sub-network reconstructing the CT images from the different geometries. Moreover, the inverse Radon transform approximation sub-network contains sinogram-domain filtering module, back-projection module, and image-domain filtering module, which maps the Radon transformation to the network. Finally, the proposed DM-MG can reconstruct the CT images from the different imaging geometries and dose levels simultaneously. In this work, the CT data from 2016 NIHAAPM-Mayo Clinic Low Dose CT Grand Challenge are used to validate and evaluate the reconstruction performance of the proposed DM-MG. Experimental results demonstrate that the presented DM-MG method can produce high-quality CT images from multiple geometries simultaneously and outperform the competing method in the quantitative assessments and visual inspection show that the dual-domain modulated model we propose provides better reconstruction of low-dose CT images under different imaging geometries compared to other modulated models.

Deep learning technology has been utilized in computed tomography, but, it needs centralized dataset to train the neural networks. To solve it, federated learning has been proposed, which collaborates the data from different local medical institutions with privacy-preserving decentralized strategy. However, lots of unpaired data is not included in the local models training and directly aggregating the parameters would degrade the performance of the updated global model. In order to deal with the issues, we present a semi-supervised and semi-centralized federated learning method to promote the performance of the learned global model. Specifically, each local model is trained with an unsupervised strategy locally at a fixed round. After that, the parameters of the local models are shared to aggregate on the server to update the global model. Then, the global model is further trained with a standard dataset, which contains well paired training samples to stabilize and standardize the global model. Finally, the global model is distributed to local models for the next training step. For shorten, we call the presented federated learning method as “3SC-FL”. Experiments demonstrate the presented 3SC-FL outperforms the compared methods, qualitatively and quantitatively.

Next generation X-ray computed tomography, based on photon-counting detectors, is now clinically available. These new detectors come with the promise of higher contrast-to-noise ratio and spatial resolution and improved low-dose imaging. However, the multi-bin nature of photon-counting detectors renders the image reconstruction problem more difficult. Common approaches, such as the two-step projection-based approach, may result in material basis images with an excessive degree of noise, which limits the clinical usefulness of the images. One possible solution is to “assist” the conventional image reconstruction by post-processing the reconstructed images using deep learning. Such networks are often trained using some pixel-wise loss, such as the mean squared error. This low-level per-pixel comparison is known to lead to over-smoothing and a loss of fine-grained details that are important to the perceptual quality and clinical usefulness of the image. In this abstract, we propose to tackle this issue by including an adversarial loss based on the Wasserstein generative adversarial network with gradient penalty. The adversarial loss will encourage the distribution of the processed images to be similar to that of the ground truth. This helps prevent over-smoothing and ensures that the ground truth texture is well preserved. In particular, we train a version of the UNet using a combination of the mean absolute error and an adversarial loss to correct for noise in the material basis images. We demonstrate that the proposed method can produce denoised virtual monoenergetic images, with realistic texture, at a range of energy levels.

Computed tomography (CT) is a widely used medical imaging modality which is capable of displaying the fine details of human body. In clinics, the CT images need to highlight different desired details or structures with different filter kernels and different display windows. To achieve this goal, in this work, we proposed a deep learning based ”All-in-One” (DAIO) combined visualization strategy for high-performance disease screening in the disease screening task. Specifically, the presented DAIO method takes into consideration of both kernel conversion and display window mapping in the deep learning network. First, the sharp kernel, smooth kernel reconstructed images and lung mask are collected for network training. Then, the structure is adaptively transferred to the kernel style through local kernel conversion to make the image have higher diagnostic value. Finally, the dynamic range of the image is compressed to a limited gray level by the mapping operator based on the traditional window settings. Moreover, to promote the structure details enhancement, we introduce a weighted mean filtering loss function. In the experiment, nine of the ten full dose patients cases from the Mayo clinic dataset are utilized to train the presented DAIO method, and one patient case from the Mayo clinic dataset are used for test. Results shows that the proposed DAIO method can merge multiple kernels and multiple window settings into a single one for the disease screening.

Supervised deep learning (DL) methods have been widely developed to remove noise-induced artifacts and promote image quality in the low-dose CT imaging task via good mapping capabilities. These supervised DL methods are usually trained based on a large amount of low- and normal-dose sinogram/image pairs and the reconstruction performance of such supervised DL methods heavily depends on the quality of reference training images. In the CT imaging, it is challenging to collect lots of high quality reference training images in practice due to the risk of high radiation dose to patients. Moreover, the medium quality or even low quality reference training images (i.e., the low quality labeled data with some noise-induced artifacts) might be collected for the supervised DL networks training, which would degrade the reconstruction performance of the network. To address this issue, in this work, we propose an effective noise-conscious explicit weighting network (NEW-Net) for low-dose CT imaging wherein the CT images with noise-induced artifacts are treated as labeled data in the network training. Specifically, the proposed NEW-Net consists of two sub-networks, i.e., noise estimation sub- network, and noise-conscious weighting sub-network. The noise estimation sub-network produces the noise map from the low-quality training data to estimate the noise-conscious weights, which determines the contribution of the label data, i.e., small weights go along with the low quality label data with severe nosie-induced artifacts, and large weights go along with high quality label data with a few noise-induced artifacts. Then the estimated weights are used to condition the training data to train the noise-conscious weighting sub-network to eliminate the effects of low quality label data and promote the reconstruction performance and stability of the proposed NEW-Net method. The Mayo clinic data are utilized to validate and evaluate the reconstruction performance of the proposed NEW-Net method. And the experimental results demonstrate that the proposed NEW-Net method outperforms the other competing methods in the case of low quality training data, in terms of noise-induced artifacts and structure detail preservation both qualitatively and quantitatively.

Deep learning (DL)-based algorithms have shown promising performance in low-dose computed tomography (LDCT) and are becoming mainstream methods. These DL-based methods focus on different aspects of CT image restoration, such as noise suppression, artifacts removal, structure preservation, etc. Therefore, in this paper, we propose a bayesian ensemble learning network (BENet) that fuses several representative denoising algorithms from the denoiser pool to improve the LDCT imaging performance. Specifically, we first select four advanced CT image and natural image restoration networks, including REDCNN, FBPConvNet, HINet, and Restormer, to form the denoiser pool, which integrates the denoising capabilities of different networks. The denoiser pool is pre-trained to obtain the denoising results of each denoiser. Then, we present a bayesian neural network to predict the weight maps and variances of denoiser pool by modeling the aleatoric and epistemic uncertainties of DL. Finally, the predicted pixel-wise weight maps are used to fuse the denoising results to obtain the final reconstruction result. Qualitative and quantitative analysis results have shown that the proposed BENet can effectively boost the denoising performance and robustness of LDCT image reconstruction.

Inspired by the deep learning techniques, data-driven methods have been developed to promote image quality and material decomposition accuracy in dual energy computed tomography (DECT) imaging. Most of these data-driven DECT imaging methods exploit the image priors within large amount of training data to learn the mapping function from the noisy DECT images to the desired high-quality material images in a supervised manner. Meanwhile, these supervised DECT imaging methods only estimate the multiple material images directly from the network but the material decomposition mechanism is not included in the network, and they fail to consider the unlabeled noisy DECT images to further improve the performance. In this work, to address these issues, we propose a novel Weak-supervised learning Multi-material Decomposition Network with self-attention mechanism (WMD-Net) to estimate multiple material images from DECT images with the combination of labeled and unlabeled DECT images accurately and effectively. Specifically, in the proposed WMD-Net, the labeled DECT images are used to estimate the three material images in a supervised sub-network, and the unlabeled DECT images are used to construct the unsupervised sub-network with the benefit of material decomposition mechanism. Finally, the two sub-networks are introduced into the proposed WMD-Net method. The proposed WMD-Net method is validated and evaluated through the synthesized clinical data, and the experimental results demonstrate that the proposed WMD-Net method can estimate more accurate material images than the other competing methods in terms of noise-induced artifacts reduction and structure details preservation.

Dual Energy CT (DECT) has ability to characterize different materials and quantify the densities or proportions of different contrast agents. However, the basis images decomposition is an ill-posed problem and the traditional model-based and image-domain direct inversion methods always suffer from serious degradation of the signal-to-noise ratios (SNRs). To this issue, we propose a new strategy by combining model-based and learning-based methods, which suppresses noise in the material images after direct inverse, and design a semi-supervised framework, Adaptive Semi-supervised Learning Material Estimation Network (ASLME-Net), to balance the detail structure preservation and noise suppression when fed little paired data in training stage of the deep learning. Specifically, the ASLME-Net contains two sub-networks, i.e., supervise sub-network and unsupervised sub-network. The supervised sub-network aims at capturing key features learned by with the labeled data, and the unsupervised sub-network adaptively learns the transferred feature distribution from supervised sub-network with Kullback-Leibler (KL) divergence. Experiment shows that the presented method can suppress the noise propagation in decomposition and yield qualitatively and quantitatively accurate results during the process of material decomposition. To this issue, we propose a new strategy by combining model-based and learning-based methods, which suppresses noise in the material images after direct inverse, and design a semi-supervised framework, Adaptive Semi-supervised Learning Material Estimation Network (ASLME-Net), to balance the detail structure preservation and noise suppression when fed little paired data in training stage of the deep learning.

Dual-energy computed tomography (DECT) is a promising technology that has shown a number of clinical advantages over conventional X-ray CT, such as improved material identification, artifact suppression, etc. For proton therapy treatment planning, besides material-selective images, maps of effective atomic number (Z) and relative electron density to that of water (ρ_e) can also be achieved and further employed to improve stopping power ratio accuracy and reduce range uncertainty. In this work, we propose a one-step iterative estimation method, which employs multi-domain gradient L₀-norm minimization, for Z and ρ_e maps reconstruction. The algorithm was implemented on GPU to accelerate the predictive procedure and to support potential real-time adaptive treatment planning. The performance of the proposed method is demonstrated via both phantom and patient studies.

In photon counting detectors (PCDs), electric pulses induced by two or more x-ray photons can pileup when their temporal separation is less than the detector deadtime. Pulse pileups degrade the quantitative accuracy of PCD projection and CT images. In contrast, conventional energy integrating detectors (EIDs) that integrate x-ray induced electric charge are linear with x-ray flux without suffering from pileup losses even at high flux levels. This work presents a novel PCD readout circuit design that leverages the fact that the total induced charge is linearly related to the true input flux to correct for pileup-induced count losses in PCDs. The method only requires an inexpensive upgrade to the existing PCD circuits to record the accumulated charges. After collecting PCD counts and the integrated charges, a lookup table was used to estimate the true photon events. Proof-of-concept experiments were performed with a CdTe-based PCD, and an EID was used to emulate the proposed charge summing component of the readout circuit. With the proposed correction, PCD counts became linear with input flux, even under severe pulse pileup conditions where the PCD is paralyzed and the uncorrected counts start to decrease with increasing flux. Without the correction, post-log sinograms for the test objects were severely overestimated; with the proposed correction, the sinograms accurately represent the true radiological path lengths of the objects. Lastly, no impact on spatial resolution was observed after the proposed correction in images of a spatial resolution test pattern.

Multi-energy CT can be used to decompose CT images into different material bases to allow for the quantification of materials. Three-material decomposition is desirable for clinical applications such as K-edge imaging. However, three-material decomposition requires measurements at 3 or more energies or additional assumptions such as volume conservation. This work concerns the following questions: 1) For photon counting detector (PCD) CT with only two energy bins, can the joint use of the PCD and kV-switching improve the performance of three-material decomposition for K-edge imaging? 2) If the PCD energy threshold must be fixed for both kV levels, how can its optimal value be pre-determined for a given K-edge imaging task? 3) While in conventional kV-switching multi-energy CT the separation of the two kV levels needs to be as wide as achievable, is this still the case when kV-switching is jointly used with a dual-bin PCD? These questions were answered by performing Cram´er-Rao lower bound (CRLB) analysis for dual-bin PCD-CT with kV-switching to optimize the PCD threshold and kV levels for the decomposition of water, iodine, and gadolinium bases using a previously modeled CdTe PCD energy response function. Experimental dual-bin PCD-CT images were acquired with kV-switching, and without kV-switching but with the volume conservation constraint, using the optimal acquisition parameters obtained from the CRLB analysis. Dual-bin PCD-CT with kV-switching allowed separation of all three basis materials with improved quantitative accuracy over dual-bin PCD-CT without kV-switching, which also suffered from poorer material separation and residual or missing signals from other materials.

Modern CT enables fast volumetric helical acquisition with collimations up to 80 mm. These fast acquisitions are desirable to reduce patient time in the scanner and thus the likelihood of motion during the scan, especially for pediatric patients. Traditional approximate analytic reconstruction methods produce cone-beam artifacts in wide-coverage helical scan modes. These artifacts limit the clinical utility of wide-coverage acquisitions, as lower collimations are often selected to reduce the influence of these artifacts. Here, we develop and demonstrate the merits of a fast and effective analytic method for helical cone-beam artifact reduction (CBAR) which is suitable for clinical CT reconstruction. The hybrid reconstruction method described here reconstructs two image volumes: one with lower image noise and one with lower levels of cone beam artifact. The images are then combined using a two-dimensional Fourier blending approach. We demonstrate the methods effectiveness using phantoms (uniform and anthropomorphic) and clinical image data from head and neck exams in which these artifacts are most visible. When compared with traditional weighting, helical CBAR exhibited: a quantitative reduction in cone-beam artifacts, comparable noise values in uniform water phantoms, and a Likert-score increase from 2.8/5 to 4.1/5 (over a range of neurological CT scan types, N = 9). In conclusion, the frequency-blending hybrid reconstruction for helical CBAR has been demonstrated to be both fast and effective, providing higher diagnostic confidence when reading images from wide-coverage acquisitions and potentially enabling more frequent use of shorter-duration, wide-coverage helical scan modes such as 80 mm collimation.

Radiography and Fluoroscopy (R&F) modalities are used typically for diagnostic procedures such as GI studies using a contrast agent like barium or Gastografin and vascular studies using iodine. In older analog systems using an image intensifier for fluoro, spot image acquisition at higher doses is accomplished using a light sensor to measure the brightness of the light produced by the image intensifier. Modern R&F systems employing a flat-panel detector use an Automatic Exposure Control (AEC) or ion chamber designed for radiographic use to terminate the exposure. Typically, the AECs are set to their highest sensitivity producing spot images with about 30x the dose of a frame of fluoro imaging. Given the limitations of dynamic flat-panel detector technology, a much lower spot dose of only about 10x higher than fluoroscopic dose is more appropriate. These limitations include exposure time windows and dynamic range. The sensitivity of ion chambers used for AEC is determined fundamentally by their active volume and the gain of the preamplifier. Most radiographic AECs use three to five fields whose outputs are usually averaged. An AEC with a single, much larger ion chamber field provides a 3x improvement in sensitivity without changing the gain of the preamplifier compared to existing AECs used for RAD. Such an AEC has proven to lower the Spot image dose by a factor of three with improved contrast detail in R&F applications without the associated losses in signal to noise ratio associated with raising the gain of the preamplifier.

This study explores reverberant shear wave elastography to create accurate magnetic resonance elastograms. The reverberant elastography technique utilizes the complex wave field originating from multiple point sources or reflected from various angles and superimposed with each other. The study was conducted on a calibrated brain phantom. Results showed that reverberant elastography produced accurate elastograms with an accuracy range of 84-97% and contrast-to-noise ratios of 24 dB, compared to an accuracy range of 86-97.7% and contrast-to-noise ratios of 25 dB for the established subzone inversion method.

Convolutional neural network (CNN)-based material decomposition has the potential to improve image quality (visual appearance) and quantitative accuracy of material maps. Most methods use deterministic CNNs with mean-square-error loss to provide point-estimates of mass densities. Point estimates can be over-confident as the reliability of CNNs is frequently compromised by bias and two major uncertainties – data and model uncertainties originating from noise in inputs and train-test data dissimilarity, respectively. Also, mean-square-error lacks explicit control of uncertainty and bias. To tackle these problems, a Bayesian dual-task CNN (BDT-CNN) with explicit penalization of uncertainty and bias was developed. It is a probabilistic CNN that concurrently conducts material classification and quantification and allows for pixel-wise modeling of bias, data uncertainty, and model uncertainty. CNN was trained with images of physical and simulated tissue-mimicking inserts at varying mass densities. Hydroxyapatite (nominal density 400mg/cc) and blood (nominal density 1095mg/cc) inserts were placed in different-sized body phantoms (30 – 45cm) and used to evaluate mean-absolute-bias (MAB) in predicted mass densities across different images at routine- and half-routine-dose. Patient CT exams were collected to assess generalizability of BDT-CNN in the presence of anatomical background. Noise insertion was used to simulate patient exams at half- and quarter-routine-dose. The deterministic dual-task CNN was used as baseline. In phantoms, BDT-CNN improved consistency of insert delineation, especially edges, and reduced overall bias (average MAB for hydroxyapatite: BDT-CNN 5.4mgHA/cc, baseline 11.0mgHA/cc and blood: BDT-CNN 8.9mgBlood/cc, baseline 14.0mgBlood/cc). In patient images, BDT-CNN improved detail preservation, lesion conspicuity, and structural consistency across different dose levels.

The RECIST criteria are used in computed tomography (CT) imaging to assess changes in tumour burden induced by cancer therapeutics throughout treatment. One of its requirements is frequent measurement of lesion diameters , which is often time consuming for clinicians. We aimed to study clinician-interactive AI, defined as deep learning models that use image annotations as input to assist in radiological measurements. Two annotation types are compared in their enhancement of predictive capabilities: mouse clicks in the tumour region, and bounding boxes surrounding lesions. The model architectures compared in this study are the U-Net, V-Net, AH-Net, and SegRes-Net. Models were trained and tested using a non-small cell lung cancer dataset from the cancer imaging archive (TCIA) consisting of CT scans and corresponding gold-standard lesion segmentations inferred from PET/CT scans. Mouse clicks and bounding boxes, representing clinician input, were artificially generated. The absolute percent error between predicted and ground truth diameters was computed for each model architecture. Bounding box annotations yielded mean absolute percent errors of 4.9 ± 2.1 %, 7.8 ± 3.4 %, 5.6 ± 2.4 % and 5.6 ± 2.3 %, respectively, whereas models using clicks annotations yielded 17.0 ± 7.9 %, 19.8 ± 9.3 %, 21.4 ± 10.9 % and 18.1 ± 7.9%. The corresponding mean dice scores across all model architectures were 0.883 ± 0.004 and 0.760 ± 0.012 for bounding box and click annotations respectively. Models were then implemented in an AI pipeline for clinical use at the BC cancer agency using the Ascinta software package; click annotations yielded qualitatively better results than bounding box annotations.

For the detection of very small objects, high resolution detectors are expected to provide higher dose efficiency. We assessed this impact of increased resolution on a clinical photon counting detector CT (PCD-CT) by comparing its detectability in high resolution and standard resolution (with 2x2 binning and larger focal spot) modes. A 50𝜇𝑚-thin metal wire was placed in a thorax phantom and scanned in both modes at three exposure levels (12, 15, and 18 mAs); acquired data were reconstructed with three reconstruction kernels (Br40, Br68, and Br76, from smooth to sharp). A scanning nonprewhitening model observer searched for the wire location within each slice independently. Detection performance was quantified as area under the exponential transform of the free response ROC curve. The high-resolution mode had the mean AUCs at 18 mAs of 0.45, 0.49, and 0.65 for Br40, Br68, and Br76, respectively, which were 2 times, 3.6 times, and 4.6 times those of the standard resolution mode. The high-resolution mode achieved greater AUC at 12 mAs than the standard resolution mode at 18 mAs for every reconstruction kernel, but improvements were larger at sharper kernels. The results are consistent with the greater suppression of noise aliasing expected at higher frequencies with high resolution CT. This work illustrates that PCD-CT can provide large dose efficiency gains for detection tasks of small, high contrast lesions.

Dual energy (DE) CT is on the path to dramatically improve the impact of CT on patient healthcare. We are particularly interested in accurate estimation of tissue attenuation values, a key potential of DECT yet to be achieved. In many implementations of DECT, the two scans can be decomposed in the projection domain before image reconstruction. This decomposition amounts to solving a pair of non-linear equations for each measurement line. A computationally efficient approach addressing this problem is to directly express the solution with a polynomial fit. Most often, a cubic order is deemed sufficient, and the polynomial coefficients are found with calibration scans through combinations of positive lengths of the basis materials. Alternatively, when the energy response of the measurements is known, calibration scans are not needed, and both negative and positive material lengths can be considered. Negative material lengths can be useful to account for quantum noise and partial volume effects, and to enrich the set of tissues that can be exactly represented by the basis materials. In this work, we study the accuracy of the polynomial fitting assuming that the measurements have been corrected for scatter and the energy responses are known, using a data processing pipeline we recently published. Experimental results based on CT scans of the ACR phantom show that negative lengths are not uncommon, that fitting based on positive lengths therefore inherently involve unsafe extrapolation, and that quintic polynomial fitting with negative and positive material lengths yield robust and accurate results.

Imaging is often a first-line method for diagnostics and treatment. Radiological workflows increasingly mine medical images for quantifiable features. Variability in device/vendor, acquisition protocol, data processing, etc., can dramatically affect quantitative measures, including radiomics. We recently developed a method (PixelPrint) for 3D-printing lifelike computed tomography (CT) lung phantoms, paving the way for future diagnostic imaging standardization. PixelPrint generates phantoms with accurate attenuation profiles and textures by directly translating clinical images into printer instructions that control density on a voxel-by-voxel basis. The present study introduces a library of 3D printed lung phantoms covering a wide range of lung diseases, including usual interstitial pneumonia with advanced fibrosis, chronic hypersensitivity pneumonitis, secondary tuberculosis, cystic fibrosis, Kaposi sarcoma, and pulmonary edema. CT images of the patient-based phantom are qualitatively comparable to original CT images, both in texture, resolution and contrast levels allowing for clear visualization of even subtle imaging abnormalities. The variety of cases chosen for printing include both benign and malignant pathology causing a variety of alveolar and advanced interstitial abnormalities, both clearly visualized on the phantoms. A comparison of regions of interest revealed differences in attenuation below 6 HU. Identical features on the patient and the phantom have a high degree of geometrical correlation, with differences smaller than the intrinsic spatial resolution of the scans. Using PixelPrint, it is possible to generate CT phantoms that accurately represent different pulmonary diseases and their characteristic imaging features.

Designing a synthetic crown is a time-consuming, inconsistent, and labor-intensive process. In this work, we present a fully automatic method that not only learns human design dental crowns, but also improves the consistency, functionality, and esthetic of the crowns. Following success in point cloud completion using the transformer-based network, we tackle the problem of the crown generation as a point-cloud completion around a prepared tooth. To this end, we use a geometry-aware transformer to generate dental crowns. Our main contribution is to add a margin line information to the network, as the accuracy of generating a precise margin line directly, determines whether the designed crown and prepared tooth are closely matched to allow appropriate adhesion. Using our ground truth crown, we can extract the margin line as a spline and sample the spline into 1000 points. We feed the obtained margin line along with two neighbor teeth of the prepared tooth and three closest teeth in the opposing jaw. We also add the margin line points to our ground truth crown to increase the resolution at the margin line. Our experimental results show an improvement in the quality of the designed crown when considering the actual context composed of the prepared tooth along with the margin line compared with a crown generated in an empty space as was done by other studies in the literature. Our code is available at : “https://github.com/Golriz-code/shellGeneration/tree/main/Shell%20Generation”

Photon-counting detectors are greatly improving the resolution and image quality in computed tomography (CT) technology. The drawback is, however, that the reconstruction becomes more challenging. This is because there is a considerable increment of the processing data due to the multiple energy bins and materials in the reconstruction analysis, as well as improved resolution. Yet efficient material decomposition and reconstruction methods tend to generate noisy images that do not completely satisfy the expected image quality. Therefore, there is a need for efficient denoising of the resulting material images. We present a new and fast denoiser that is based on a linear minimum mean square error (LMMSE) estimator. The LMMSE is very fast to compute, but not commonly used for CT image denoising, probably due to its inability to adapt the amount of denoising to different parts of the image and the difficulty to derive accurate statistical properties from the CT data. To overcome these problems we propose a model-based deep learning strategy, that is, a deep neural network that preserves an LMMSE structure (model-based), providing more robustness unseen data, as well as good interpretability to the result. In this way, the solution adapts to the anatomy in every point of the image and noise properties at that particular location. In order to asses the performance of the new method, we compare it to both to a conventional LMMSE estimator and to a “black-box” CNN in a simulation study with anthropomorphic phantoms.

This conference presentation was prepared for the Medical Imaging 2023: Physics of Medical Imaging conference at SPIE Medical Imaging 2023.

Noise power spectrum (NPS) is an indispensable component in the quantitative assessment of x-ray computed tomography (CT) image quality. In principle, an accurate experimental measurement of the NPS in CT requires many repeated CT image acquisitions under the same conditions. Repeated CT scanning of the same human subject is not feasible due to the obvious ethical and safety concerns associated with the increased radiation exposure. Alternatively, the NPS can be estimated using conventional phantom-based methods via repeated measurements. However, phantom-derived results 1) do not necessarily represent the patient-specific noise characteristics of CT images and 2) do not show local NPS concerning a very small region either. Therefore, it is highly desirable to directly estimate the local NPS directly from patient CT data without resolving to repeated measurements. In this work, a fast and highly accurate one-sample approach was proposed to enable patient-specific local NPS measurement with respect to a single pixel from a single CT data acquisition. In numerical studies, the one-sample approach was superior compared to the conventional multi-sample approach with 50,000 repeated measurements.

Laser-Compton x-ray sources have many advantages over traditional x-ray tubes for use in medical imaging due to their monoenergetic energy spectrum, tunability, high-flux, and low-dose potential. These properties can specifically be taken advantage of in the context of K-edge subtraction (KES) imaging. Previous optimization approaches are time-consuming by scanning over high-dimensional parameter spaces. Here, we show how a Bayesian optimization routine optimizes LCS source parameters in only a fraction of the computational time. Using this approach, we found a configuration that produces non-inferior image quality in KES mammography compared to a previously optimized direct energy tuning technique. Moreover, a successfully optimized implementation of scanning K-edge subtraction imaging was realized applying this Bayesian approach.

Coincidence timing calibration is fundamental to PET imaging. The electronics, cable lengths, and detector physics such as charge drift and depth dependence add to the measured time differences in coincidence sorting – increasing random rate, decreasing true rate, and degrading system performance. This work investigates the parameter selection for convex optimization (Ordinary Least Squares) for timing calibration. We test the correlation between commonly selected parameters and the experimentally measured coincidence time difference. Additionally, we test 127 nested models of a parameterized regression equation to identify the those which optimize MSE, BIC, and FWHM, respectively. In each of these models, the FWHM performance improved ~53%, though the value shifted from ~ 304 to 160 ns – far from ~ 10 ns FWHM CZT can achieve. The results point to the lack of a necessary parameter, such as trigger threshold level or temperature, or data which is too variable for the OLS optimization.

Breathing can cause blurring and artifacts in PET images, especially in the body trunk region. The blurring and artifacts can negatively impact cancer detection and response to therapy assessment. Many motion detection techniques, such as using external motion sensors or data driving methods, have been used to facilitate respiratory motion correction for PET. These methods require sophisticated gating or motion compensated image reconstruction, which are time consuming. In our work, we propose a deep learning framework based on U-Net to directly perform respiratory motion correction for whole-body PET in the image domain. Our framework was trained with the patches of the PET images without motion correction as input and the motion-corrected images using the data-driving gating (DDG) method as label. The framework also incorporated spatial information of voxels by adding additional layer of voxels’ vertical locations to the input. To evaluate our framework, we conducted 5-fold cross validations to generate the motion-corrected images for 30 subjects and compared them with the ground truth images corrected by the DDG methods. Our framework could correct the PET images in regions affected by respiratory motion. The incorporation of voxels’ spatial information could further improve the performance of motion correction. There is a great potential of our framework to perform direct respiratory motion correction in the image domain in a convenient manner.

Benchtop experiments have proven the utility of a stationary computed tomography (CT) scanner for head imaging. The purpose of this study was for system control development and integration in a clinical setting for clinical use and evaluation. Software and interfaces for technologist operation of the complete system during patient scanning were also developed. A clinical imaging bed was integrated with the x-ray control system with off-the-shelf microcontrollers to drive the development of the system for clinical evaluation. The clinical imaging system is composed of three carbon nanotube (CNT) x-ray source arrays and nine strip detectors. 135 projections are acquired per slice at 120kVp, 10mA, and 2.95ms exposure per projection. Anthropomorphic phantoms have been imaged in preparation for the first patients. Reconstruction was performed using adaptive steepest descent projection onto convex set (ASD-POCs) method. Scanner performance parameters were measured. Images are evaluated by neuroradiologists. A working stationary head CT (sHCT) system has been developed for patient imaging evaluation. Image quality is sufficient for starting the observational clinical trial as shown in images of the ACR accreditation phantom and KYOTO head phantoms. This preliminary study has shown that the sHCT system is ready for patient imaging studies. Clinical utility will be assessed in a patient study with patients with prior head trauma.