This PDF file contains the front matter associated with SPIE Proceedings Volume 12304, including the Title Page, Copyright information, Table of Contents, and Conference Committee lists.

Computed tomography (CT) is a foundation of modern clinical diagnostics but it presently only retrieves information from X-rays attenuation. However, it is known that micro structural texture or porosity – which is well below the spatial resolution of CT – can be revealed by grating-based dark-field imaging. Diagnostic value of this sub-resolution tissue information has been demonstrated in pre-clinical studies on small-animal disease models and recently also in a first clinical radiography system.¹ These studies show that dark-field imaging is particularly useful for early detection and staging of lung diseases. While dark-field CT is regularly realized in laboratory environment, the transfer to human scale and bringing it to clinical application poses several technical challenges. Switching from a step-and-shoot acquisition to a mode where the gantry and acquisition operate continuously as well as reducing scan times to below seconds and ensuring stability against vibrations are key concerns when it comes to the translation of the established laboratory dark-field technology to full-body medical CT. In,² we recently demonstrated the first dark-field CT implementation, which collectively solves these roadblocks and therefore is a milestone in the development of clinical CT imaging. The prototype we present allows to reconstruct the attenuation and dark-field channels of a human thorax phantom from a one second long acquisition and covers a 45 cm diameter field of view. In this work, we present how the first dark-field CT prototype works and focus particularly on the technical design, optimized design of the gratings for CT application and the first characterization of the interferometer in the rotating gantry. We discuss which steps where particularly important for the realization and where we see potential for further improvements. These results provide key insights for future dark-field CT implementations.

X-ray computed tomography (CT) has been established as a daily tool in clinical diagnostics and has been continuously refined by more recent innovations in the last years. These systems are, however, limited by fundamental constraints since they are only capable of mapping X-ray attenuation differences in the tissue. Phase-contrast and dark-field imaging provide complementary contrast, which originates from physically different interaction processes of X-rays with matter. Particularly the dark-field signal is considered to have significant diagnostic potential since it is capable to retrieve micro-structural information below the actual resolution limit of the imaging system. This was demonstrated in various laboratory setups and recently also in the fist study with human patients in a clinical radiography system based on a grating interferometer. In a recent work, we presented the first implementation of such an X-ray interferometer into a clinical CT gantry. Upscaling and adapting this technology for a rotating CT gantry involves several challenges and tradeoffs ranging from limitations in interferometer design over fast, continuous signal acquisition requirements to tolerances in applied patient dose. In this work we discuss the performance of the first clinical dark-field CT prototype. For this purpose, we present results of our phantom studies which were designed to evaluate whether and how the dark-field contrast generated by the system is capable to provide additional structural sample information. The key aspects include the possibility of quantitative imaging and a gradual approach to simulate results that come as close as possible to a real application in a human patient.

Efficient removal of solid focal tumors is a major challenge in modern medicine. Percutaneous thermal ablation is a first-line treatment for patients not fit for surgical resection or when the disease burden is low, mainly due to expedited patient recovery times, lower rates of post-operative morbidity, and reduced healthcare costs. While continuously gaining popularity, ~100,000 yearly thermal hepatic ablation procedures are currently performed without actively monitoring temperature distributions, leading to high rates of incomplete ablations, local recurrences, and damage to surrounding structures. Recent advancements in computed tomography (CT), especially spectral CT, provide promising opportunities for lowering these rates. The additional information available with spectral CT can provide the necessary capabilities to achieve accurate, reliable, on-demand, and non-invasive thermometry during ablation procedures. By taking advantage of our newly developed spectral physical density maps and their direct relation with temperature changes, we performed experiments on phantoms and ex vivo tissue to develop, evaluate, optimize, and refine a method for generating thermometry maps from spectral CT scans. Our results validate the accuracy of the spectral physical density model, allowing “whole-organ” mass quantifications that are accurate within one percent, as well as demonstrate an ability to extract temperature changes (linear correlation coefficient of 0.9781) non-invasively and in real-time.

We describe a new procedure for three-dimensional (3D) region-of-interest (ROI) reconstruction from transversely-truncated cone-beam projections acquired with a circular source trajectory. This method is an extension to 3D of the virtual fan-beam (VFB) method. It is based on a VFB formula that performs the backprojection in the acquisition geometry. Our simulation results show that the ROI reconstruction of the 3D Shepp-Logan phantom is very similar to the one obtained by the Feldkamp, Davis, Kress (FDK) algorithm without truncation. However the reconstruction of the Forbild head phantom shows artefacts which are absent from the FDK truncation-free reconstruction.

The generalized minimal residual (GMRES) algorithm is applied to image reconstruction using linear computed tomography (CT) models. The GMRES algorithm iteratively solves square, non-symmetric linear systems and it has practical application to CT when using unmatched back-projector/projector pairs and when applying preconditioning. The GMRES algorithm is demonstrated on a 3D CT image reconstruction problem where it is seen that use of unmatched projection matrices does not prevent convergence, while using an unmatched pair in the related conjugate gradients for least-squares (CGLS) algorithm leads to divergent iteration. Implementation of preconditioning using GMRES is also demonstrated.

Due to longitudinal undersampling multislice spiral computed tomography (MSCT) scans may suffer from windmill artifacts in reconstructed images. To fulfill the sampling condition and achieve double sampling in z-direction, some CT scanners use the z-flying focal spot (zFFS) technique, a hardware-based solution that effectively doubles the number of detector rows. To obtain a software-based solution we developed a convolutional neural network that is trained in a supervised manner with clinical projection raw data that were acquired with zFFS enabled. We presented this approach as the row interpolation with deep learning (RIDL) network. In this work we simplified the network architecture, extended the clinical dataset and generated an experimental synthetic dataset consisting of two-dimensional projection data. We were able to observe a reduction in windmill artifacts for both datasets used for training. Especially the synthetic dataset is very promising as we could observe a superior reduction of artifacts with this dataset.

High-precision image-guided neurosurgery – especially in the presence of brain shift – would benefit from intraoperative image quality beyond the conventional contrast-resolution limits of cone-beam CT (CBCT) for visualization of the brain parenchyma, ventricles, and intracranial hemorrhage. Deep neural networks for 3D image reconstruction offer a promising basis for noise and artifact reduction, but generalizability can be challenged in scenarios involving features previously unseen in training data. We propose a 3D deep learning reconstruction framework (termed “DL-Recon”) that integrates learning-based image synthesis with physics-based reconstruction to leverage strengths of each. A 3D conditional GAN was developed to generate synthesized CT from CBCT images. Uncertainty in the synthesis image was estimated in a spatially varying, voxel-wise manner via Monte-Carlo dropout and was shown to correlate with abnormalities or pathology not present in training data. The DL-Recon approach improves the fidelity of the resulting image by combining the synthesized image (“DL-Synthesis”) with physics-based reconstruction (filtered back-projection (FBP) or other approaches) in a manner weighted by uncertainty – i.e., drawing more from the physics-based method in regions where model uncertainty is high. The performance of image synthesis, uncertainty estimation, and DL-Recon was investigated for the first time in real CBCT images of the brain. Variable input to the synthesis network was tested – including uncorrected FBP and precorrection with a simple (constant) scatter estimate – hypothesizing the latter to improve synthesis performance. The resulting uncertainty estimation was evaluated for the first time in real anatomical features not included in training (abnormalities and brain shift). The performance of DL-Recon was evaluated in terms of image uniformity, noise, and soft-tissue contrast-to-noise ratio in comparison to DL-Synthesis and FBP with a comprehensive artifact correction framework. DL-Recon was found to leverage the strengths of the learning-based and physics-based reconstruction approaches, providing a high degree of image uniformity similar to DL-Synthesis while accurately preserving soft-tissue contrast as in artifact-corrected FBP.

Learned iterative reconstruction algorithms for inverse problems offer the flexibility to combine analytical knowledge about the problem with modules learned from data. This way, they achieve high reconstruction performance while ensuring consistency with the measured data. In computed tomography, extending such approaches from 2D fan-beam to 3D cone-beam data is challenging due to the prohibitively high GPU memory that would be needed to train such models. This paper proposes to use neural ordinary differential equations to solve the reconstruction problem in a residual formulation via numerical integration. For training, there is no need to backpropagate through several unrolled network blocks nor through the internals of the solver. Instead, the gradients are obtained very memory-efficiently in the neural ODE setting allowing for training on a single consumer graphics card. The method is able to reduce the root mean squared error by over 30% compared to the best performing classical iterative reconstruction algorithm and produces high quality cone-beam reconstructions even in a sparse view scenario.

We investigate fast silicon photomultiplier (SiPM)-based scintillation detectors for X-ray photon-counting applications, e.g., photon-counting computed tomography (CT). Such detectors may be an alternative to CdTe/CdZnTe (CZT) and Si detectors, which face challenges related to availability and cost-effective growth of detector-grade material, and detection efficiency, respectively. Here, we experimentally study energy response and count rate performance of a 1 mm × 1 mm single-pixel detector consisting of the readily available LaBr3:Ce scintillator and an ultrafast SiPM. We used three radio-isotopes and an X-ray tube for the experiments. Raw detector signals were processed by a second-order low-pass filter with a cut-off frequency fc equal to 25 MHz or 100 MHz. The detector pulse height was shown to be proportional to photon energy. We measured FWHM energy resolutions of 19.5% (fc=25 MHz) and 21.5% (fc=100 MHz) at 60 keV. The measured X-ray tube spectra showed signs of the expected features of such spectra. The best count rate performance was achieved using fc=100 MHz. In case of paralyzable-like counting and a 30 keV counting threshold, the maximum observed count rate (OCR) was 10.5 Mcps/pixel. For nonparalyzable-like counting and the same threshold, the OCR appeared to approach an asymptotic value greater than 20 Mcps/pixel. These numbers are close to those of CdTe/CZT detectors highly optimized for photon-counting CT. In conclusion, we show promising spectral X-ray photon-counting performance of an LaBr3:Ce scintillation detector with SiPM readout. Depending on the application-specific requirements, miniaturization of the pixel size may be necessary, for which we discuss potential dose-efficient implementations.

In this work we describe a new dynamic X-ray collimator that may be used to collect sparse computed tomography projection data. Data sparsity may be user-specified and controlled both angularly and radially - allowing a broad range of acquisition strategies. We consider protocols that have fully sampled projection data for a volume-of-interest with a sparsely sampled background. Model-based reconstruction methods are adapted to process the non-uniformly sampled projections. We demonstrate the ability of a CT system with this novel dynamic collimator to provide user controllable regional image quality and dose reduction in a set of phantom experiments.

X-ray tomography reconstruction requires a set of projections that provides sufficient information for the examined region. Commonly, to ensure mathematically complete reconstruction, first, a continuous curve (trajectory) that fulfils the Tuy conditions is chosen. Second, this curve is sampled based on the Nyquist-Shannon sampling theorem. This two-step approach is efficient for most standard X-ray tomography scanning scenarios. For agile X-ray tomography systems, e.g. robot-supported computed tomography systems, choosing a set of projections based on a continuous curve is often not useful. Instead, sets of projections from arbitrary views might be necessary. This work combines the Tuy-Smith condition with conclusions from the Nyquist-Shannon sampling theorem. In particular, the maximal pixel size and a requirement for the arrangement of projections are formulated depending on the smallest relevant object feature and integrated into the Tuy-Smith condition. We derive a comprehensive condition for data completeness that can assess the completeness of any set of arbitrary projections, e.g. for complex scanning scenarios with robot-supported X-ray tomography systems.

In certain CT applications such as dental CT imaging, a scanning configuration with an offset-detector is often used for extending the field of view (FOV) of the system. While data are truncated on one-side of the detector, it remains possible to accurately reconstruct an image from the truncated data collected over a full-angular range (FAR) of 360° by use of existing analytical-based algorithms such as the FDK algorithm. However, there also exist interests in scanning configurations that collect data only over limited-angular ranges (LARs) for practical considerations, and existing algorithms generally yield reconstructions with significant artifacts from LAR data collected with an offset-detector. It has been demonstrated recently that, for non-truncated data, the directional-total-variation (DTV) algorithm can reconstruct images with significantly reduced artifacts from LAR data. In this work, we developed and tailored the DTV algorithm for image reconstruction from truncated LAR data collected with a scanning configuration employing an offset-detector. We carried out a study on image reconstruction for a number of LAR scanning configurations with an offset-detector of practical interest. The study results demonstrate that the DTV algorithm can be tailored to yield, from truncated LAR data, images with significantly reduced artifacts that are observed otherwise in images obtained with existing analytical-based algorithms.

A robotic cone-beam computed tomography (CT) scanner has inherent advantages. In particular, it enables source and detector trajectories capable of extending the field-of-view (FOV) of the reconstructed CT images, where the FOV is defined as the region in the source trajectory plane for which all ray lines are acquired (at any position and direction in this plane). Previous extensions of the FOV used an offset detector or a displaced center of rotation and a single full 360° scan. However, due to limitations in the conventional range of motion inside the treatment room to avoid collisions, some systems can only perform short scans. This paper investigates a new approach to FOV extension for an existing system by adapting the FOV extension to two complementary short scans with displaced centers of rotation. We validate this approach on numerical simulations of the Forbild thorax phantom.

We propose a consistency-based material decomposition algorithm. The method is free from any calibration procedure. The inverse spectral mixing model is approximated by a polynomial whose indeterminates are the raw-data values and whose coefficients are estimated by minimizing a consistency-based cost function. The consistency is in both the material sinograms and their mono-energetic combination. A small a priori on the object is incorporated in the minimization problem as a constraint. The method was evaluated on dual-energy simulations of a numerical phantom made of water and bone.

Developing novel contrast agents for multi-energy photon-counting (PC)CT will require a clear translation pathway from preclinical validation to clinical applications. To begin this development, we have used a clinical PCCT scanner (Siemens NAEOTOM Alpha) to study the spectral separation of a few contrast elements (Iodine, I; Gadolinium, Gd; Hafnium, Hf; Tantalum, Ta; Bismuth, Bi; Calcium, Ca) with currently available scanning protocols (fixed: 120 kVp, 20 and 65 keV thresholds). We also explored the capabilities of clinical and preclinical PCCT to image mice with sarcoma tumors injected with nanoparticles (NP). Our results indicate that Ta or Hf are complementary to I or Gd, providing excellent spectral separation for future multi-agent studies. Based on preclinical PCCT with four energy thresholds, we also conclude that additional energy thresholds will benefit clinical PCCT. Furthermore, we demonstrate the role that multi-channel denoising and reconstruction algorithms will play in pushing the bounds of spatial and spectral resolution with clinical PCCT. Performing co-clinical research will facilitate the translation of novel imaging algorithms and NP contrast agents for PCCT

Reproducibility of CT numbers represents an ongoing challenge, especially in clinical applications where exams are performed to assess disease progression or response to therapy. Dual energy CT (DECT) offers an opportunity for improved image quantification, but also presents unique issues. This work presents an initial phantom study that investigates reproducibility over time, across different scanners, and with regard to positioning of the phantom. Both an image domain and a projection data domain material decomposition method are used to create mono-energetic images from DECT data. The scanned object is the ACR CT accreditation phantom. Images were evaluated for reproducibility both inside the phantom inserts of module A as well as in regions between the inserts, where image artifacts are frequently visualized. The results demonstrate that artifacts are worse for off-centered positions. They also demonstrate that the data-based material decomposition provides comparable HU numbers within the inserts of interest like the image-based method, but provides substantially less artifacts and less HU variability in regions surrounding the inserts across the different phantom positions.

Cardiac CT is a useful tool for cardiovascular diagnostics that offers different acquisition modes, each with its advantages. The development of direct converting detector technology has resulted in the clinical translation of dual-source photon-counting CT. This takes advantage of the improved image quality at high heart rates from dual-source CT while making available spectral results for more precise material characterization and quantification. To evaluate the stability of spectral results among different acquisition modes and heart rates, a cardiac motion phantom with a rod mimicking a 50% coronary stenosis was scanned with a dual-source photon-counting CT in three different acquisition modes (retrospective dual-source spiral, prospective dual-source step-and-shoot, dual-source flash spiral) and at different heart rates (60, 80, 100 bpm). Dice scores of stenosed regions relative to a static scan, eccentricity of non-stenosed regions, full width half max, and normalized area under the curve of line profiles were calculated for iodine density maps, and virtual mono-energetic images at 40 and 70 keV. Dice scores and eccentricity were consistent and not significantly affected by acquisition mode or heart rate for spectral results. Full width half max and normalized area under the curve similarly illustrated minor differences between acquisition modes and heart rates. The consistency in these metrics demonstrate preserved image structure and allows for the use of spectral results with high confidence. Dual-source photon-counting CT will enable cardiovascular diagnostics with better material characterization and differentiation.

Due to the smaller detector pixels in photon-counting CT, coarse anti-scatter grids are used. This may lead to high frequencies in the scattered radiation and therefore moiré artifacts in the reconstructed images can occur. It has been shown that deep convolutional neural networks are very effective to correct scatter artifacts in clinical CT. In this work we present an adapted version of the deep scatter estimation (DSE) to correct for the high frequency artifacts effectively. With the use of DSE the mean absolute error of the scatter artifacts is reduced from about 8 HU to under 1 HU. At the same time the moiré artifacts can be prevented and additional post-processing in the image can be avoided.

Metallic objects in the volume of a CBCT system can cause various artifacts after image reconstruction such as bright and dark streaks, local distortions of CT values and shadowing. In the intraoperative setting, this drastically reduces clinical value and can harden decision making. Most existing approaches to reduce such artifacts rely on a threshold-based metal segmentation in the reconstruction domain, which is prone to failure; especially in cases with extreme artifacts. Faulty metal masks impair these inpainting-based MAR methods and at times even worsen image quality by introducing secondary artifacts. In this work, a novel neural network topology is proposed to segment metallic objects in CBCT reconstruction domain by leveraging information of the given raw projection images. A reconstruction operator is embedded into this architecture, which enables the model to yield projection and reconstruction domain information during end-to-end training. This cross-domain approach is compared to the self-configuring segmentation method ”nnUNet”, which predicts the three dimensional metal masks directly from the artifact corrupted reconstruction. To provide a baseline, a segmentation using a global dice-optimal threshold is determined. Segmentation results on simulated data confirmed by five-fold cross validation show that the cross-domain network yields a mean dice coefficient of 0.87 ± 0.05 at a prediction time of 4s per volume. The reference method achieves 0.86 ± 0.03 in 43s, whereas the optimal similarity using a threshold averages to 0.45 ± 0.22.

Sparse-view computed tomography (CT) becomes a major concern in the medical imaging field due to its reduced X-ray radiation dose. Recently, various convolutional neural network (CNN)-based approaches have been proposed, requiring the pairs of full and sparse-view CT images for network training. However, these paired data acquisition is impractical or difficult in clinical practice. To handle this problem, we propose the weakly-supervised learning for streak artifact reduction with unpaired sparse-view CT data. For CNN training dataset, we generate the pairs of input and target images from the given sparse-view CT data. Then, we iteratively apply the trained network to given sparse-view CT images and acquire the prior images. As the success factor of our novel framework, we estimate the original streak artifacts in the given sparse-view CT images from the prior images and subtract the estimated streak artifacts from the given sparse-view CT images. As a result, the proposed method has the best performance of lesion detection compared to the other methods.

Sparse view sampling is one of the effective ways to reduce radiation dose in CT imaging. However, artifacts and noise in sparse-view filtered back projection reconstructed CT images are obvious that should be removed effectively to maintain diagnostic accuracy. In this paper, we propose a novel sparse-view CT reconstruction framework, which integrates the projection-to-image and image-to-projection mappings to build a dual domain closed-loop learning network. For simplicity, the proposed framework is termed a closed-loop learning reconstruction network (CLRrcon). Specifically, the primal mapping (i.e., projection-to-image mapping) contains a projection domain network, a backward projection module, and an image domain network. The dual mapping (i.e., image-to-projection mapping) contains an image domain network and a forward projection module. All modules are trained simultaneously during the network training stage, and only the first mapping is used in the network inference stage. It should be noted that both the inference time and hardware requirements do not increase compared with traditional hybrid domain networks. Experiments on low-dose CT data demonstrate the proposed CLRecon model can obtain promising reconstruction results in terms of edge preservation, texture recovery, and reconstruction accuracy in the sparse-view CT reconstruction task.

In sparse X-ray Computed Tomography, the radiation dose to the patient is lowered by measuring fewer projection views compared to a standard protocol. In this work we investigate a hybrid approach combining shearlet representation with deep learning for reconstruction of sparse-view X-ray computed tomography. The proposed method is hybrid in that it reconstructs the parts that can provably be retrieved by utilizing a model-based approach, and it in-paints the parts that provably cannot through a learning-based approach. In doing so, we attempt to benefit from the best aspects of model- and learning-based methods. We demonstrate first promising results on publicly available data.

In this paper we examine whether increasing the rank of a tomographic system matrix by slightly changing the measurement lines might improve the reconstructed image quality in region-of-interest (ROI) tomography. The image quality is assessed by the size of the reconstructible region and by the reconstructed noise of the pixels in that region. A theorem is presented to specify cases when changing elements in the system matrix leads to solving new pixels. Numerical simulations were performed on several modest sized examples to measure the bias and variance of the reconstructed images using system matrices with the same size but different rank. The results show that while increasing the rank might help to solve new pixels, it can also lead to increases in the pixel variances of the initial ROI.

Hemorrhagic stroke accounts for up to 20% of all stroke cases, and requires a treatment pathway drastically different to ischemic stroke. Prompt triage is therefore crucial and often only attainable with neuroimaging for intracranial hemorrhage (ICH) evaluation, for which MDCT is the frontline modality. Availability of ICH dedicated imaging in the pre-hospital setting, with portable CT systems, would facilitate early ICH diagnosis. However, current CT or cone-beam CT (CBCT) approaches often use conventional x-ray sources mounted on a rotating gantry limiting their minimum weight and footprint. Recent advances on cold-cathode, compact X-ray sources, based on carbon nanotube (CNT) technology, enable the development of ultra-compact designs based on source-array arrangements on stationary configurations. However, such geometrical arrangements show limited angular sampling, and sparse, non-stationary, volume sampling. In this work we present first investigation of geometric configuration and effects of 3D sampling pertinent to the task of ICH detection on an ultra-portable stationary CBCT for ICH imaging. The baseline configuration included 31 CNT sources on a curved array illuminating a curved panel detector (871 mm length), on a compact geometrical configuration (SDD = 690 mm). Metrics of sampling completeness, sampling density, and MTF shape and band-width integral were explored for configurations varying in source angular span (30°-170°), source array and detector curvature radius (250 mm to flat), use of 2D matrix source arrangements, and multi-acquisition protocols. The results show that sufficient sampling and resolution can be achieved with a combination of moderate curvature (~450 mm radius) of the source array and detectors, with better sampling properties for approximately matched curvature radii (up to 30% BWI improvement). Improved image quality was demonstrated with configurations featuring matrix source arrangements in combination with multi-acquisition protocols (around a 6% of improvement in sampling completeness).

The goal of this study is to understand how the normalized glandular dose coefficient (𝐷𝑔𝑁𝐶𝑇) varies with projection angle in dedicated cone-beam breast computed tomography (CBBCT). Seventy-five CBBCT clinical datasets from a research database were used for this study. All samples were segmented into skin, adipose and fibroglandular tissues. The segmented volumes were used in a Monte Carlo simulation package (GATE 8.0) to estimate the radiation dose at 10 angles in a full scan. An analytical model is proposed, and this model predicted that the angular 𝐷𝑔𝑁𝐶𝑇 follows a sine wave and the maximum is related to the center of geometry of the fibroglandular tissue (COGf ). The angular 𝐷𝑔𝑁𝐶𝑇 from Monte Carlo simulations was consistent with our model and follows a sine wave with amplitude of 0.0376. The maximum of the wave occurs when the x-ray source is approximately at head position, which is consistent with our model. Our results indicate that the higher angular 𝐷𝑔𝑁𝐶𝑇 occurs when the x-ray source is superior to the breast. This suggests using a x-ray source trajectory inferior to the breast for short-scan CBBCT design.

Quantitative imaging biomarkers (QIBs) hold enormous potential to improve the efficiency of clinical trials that use standard-of-care CT imaging. Examples of QIBs include size, shape, intensity histogram characteristics, texture, radiomics, and more. There is, however, a well-recognized gap between discovery and the translation to practice of QIBs, which is driven in part by concerns about their repeatability and reproducibility in the diverse clinical environment. Our goal is to characterize QIB repeatability and reproducibility by using virtual imaging clinical trials (VICTs) to simulate the full data pathway. We start by estimating the probability distribution functions (PDFs) for patient-, disease-, treatment- , and imaging-related sources of variability. These are used to forward-model sinograms that are reconstructed and then analyzed by the QIB under evaluation in a virtual imaging pipeline. By repeatedly sampling from the variability PDFs, estimates of the bias, variance, repeatability and reproducibility of the QIB can be generated by comparison with the known ground truth. These estimates of QIB performance can be used as evidence of the utility of QIBs in clinical trials of new therapies.

Long lasting efforts have been made to reduce radiation dose and thus the potential radiation risk to the patient for CT acquisitions without severe deterioration of image quality. To this end, different reconstruction and noise reduction algorithms have been developed, many of which are based on iterative reconstruction techniques, incorporating prior knowledge in the image domain. Recently, deep learning-based methods have shown impressive performance, outperforming many of the previously proposed CT denoising approaches both visually and quantitatively. However, with most neural networks being black boxes they remain notoriously difficult to interpret and concerns about the robustness and safety of such denoising methods have been raised. In this work we want to lay the fundamentals for a post-hoc interpretation of existing CT denoising networks by reconstructing their invariances.

The rapid development of deep-learning methods in medical imaging has called for an analysis method suitable for non-linear and data-dependent algorithms. In this work, we investigate a local linearity analysis where a complex neural network can be represented as piecewise linear systems. We recognize that a large number of neural networks consists of alternating linear layers and rectified linear unit (ReLU) activations, and are therefore strictly piecewise linear. We investigated the extent of these locally linear regions by gradually adding perturbations to an operating point. For this work, we explored perturbations based on image features of interest, including lesion contrast, background, and additive noise. We then developed strategies to extend these strictly locally linear regions to include neighboring linear regions with similar gradients. Using these approximately linear regions, we applied singular value decomposition (SVD) analysis to each local linear system to investigate and explain the overall nonlinear and data-dependent behaviors of neural networks. The analysis was applied to an example CT denoising algorithm trained on thorax CT scans. We observed that the strictly local linear regions are highly sensitive to small signal perturbations. Over a range of lesion contrast from 0.007 to 0.04 mm⁻¹, there is a total of 33992 linear regions. The Jacobians are also shift-variant. However, the Jacobians of neighboring linear regions are very similar. By combining linear regions with similar Jacobians, we narrowed down the number of approximately linear regions to four over lesion contrast from 0.001 to 0.08 mm⁻¹. The SVD analysis to different linear regions revealed denoising behavior that is highly dependent on the background intensity. Analysis further identified greater amount of noise reduction in uniform regions compared to lesion edges. In summary, the local linearity analysis framework we proposed has the potential for us to better characterize and interpret the non-linear and data-dependent behaviors of neural networks.

Recent studies have proposed to optimize deep learning-based CT reconstruction methods for signal detectability performance. However, obtaining objective measures of signal detectability performance of the trained reconstruction networks is challenging due to the non-linear nature of the reconstruction. We propose a simple evaluation metric based on the model observer framework. The metric is based on the performance of a specific linear observer on signal-known-exactly/background-known-exactly task. The linear observer uses the signal Laplacian as a template, which we hypothesize is a better proxy for a human model observer than the ideal/Hotelling observer. We illustrate that the proposed metric can be used to select training hyper-parameters for a CNN-model used to reconstruct synthetic sparse-view breast CT data.

Dual-energy computed tomography (DECT) imaging plays an important role in clinical diagnosis applications due to its material decomposition capability. However, in the cases of low-dose DECT imaging and ill-conditioned issue, the direct decomposed material images from DECT images would suffer from severe noise-induced artifacts, leading to low quality and accuracy. In this paper, we propose a self-supervised Nonlocal Spectral Similarity-induced Decomposition Network (NSSD-Net) to produce decomposed material images with high quality and accuracy in the low-dose DECT imaging. Specifically, we first build the model-driven iterative decomposition model and optimize the objective function by the iterative shrinkage-thresholding algorithm (ISTA) with the convolutional neural network. Considering the intrinsic characteristics information (i.e., structural similarity and spectral correlation) underlying DECT images, which can be used as the prior information to improve the accuracy of the decomposed material images, we construct the nonlocal spectral similarity-based cost function by using the prior information and incorporating it into the iterative decomposition network to guarantee stability. The proposed NSSD-Net method was validated and evaluated with real clinical data. Experimental results showed that the presented NSSD-Net method outperforms the other competing methods in terms of noise-induced artifacts reduction and decomposition accuracy.

The maximum likelihood (ML) principle has been a gold standard for estimating basis line-integrals due to the optimal statistical property. However, the estimates are sensitive to noise from large attenuations or low dose levels. One may apply filtering in the estimated basis sinograms or model-based iterative reconstruction. Both methods effectively reduce noise, but the degraded spatial resolution is a concern. In this study, we propose a likelihood-based bilateral filter (LBF) for the estimated basis sinograms to reduce noise while preserving spatial resolution. It is a post-processing filtration applied to the ML-based basis line-integrals, the estimates with a high noise level but minimal degradation of spatial resolution. The proposed filter considers likelihood in neighbours instead of weighting by pixel values as in the original bilateral filtration. Two-material decomposition (water and bone) results demonstrate that the proposed method shows improved noise-to-spatial resolution tendency compared to conventional methods.

Polychromatic reconstruction is a promising technique for quantitative cone-beam computed tomography in radiation therapy. In this study, we have implemented polychromatic forward projection into our reconstruction framework to directly reconstruct relative electron density volumes without the need for additional HU calibration. The underlying spectral model takes beam hardening into account by design. Thereby this extended reconstruction framework is a natural step in the direction of spectral imaging, albeit without any hardware modifications. Reconstructed relative electron density volumes from phantom scans show sufficiently good agreement with ground truth for photon dose calculation; relative errors for most inserts are below 3%. We also demonstrate beam hardening artifact reduction in virtual monoenergetic images obtained from polychromatic reconstruction as compared to an established iterative reconstruction using water-based correction. Similarly, polychromatic reconstruction shows potential for mitigating metal artifacts in a clinical scan acquired for a patient with bilateral hip implants.

We investigate the feasibility of bone marrow edema (BME) detection using a kV-switching Dual-Energy (DE) Cone-Beam CT (CBCT) protocol. This task is challenging due to unmatched X-ray paths in the low-energy (LE) and high-energy (HE) spectral channels, CBCT non-idealities such as X-ray scatter, and narrow spectral separation between fat (bone marrow) and water (BME). We propose a comprehensive DE decomposition framework consisting of projection interpolation onto matching LE and HE view angles, fast Monte Carlo scatter correction with low number of tracked photons and Gaussian denoising, and two-stage three-material decompositions involving two-material (fat-Aluminum) Projection-Domain Decomposition (PDD) followed by image-domain three-material (fat-water-bone) base-change. Performance in BME detection was evaluated in simulations and experiments emulating a kV-switching CBCT wrist imaging protocol on a robotic x-ray system with 60 kV LE beam, 120 kV HE beam, and 0.5° angular shift between the LE and HE views. Cubic B-spline interpolation was found to be adequate to resample HE and LE projections of a wrist onto common view angles required by PDD. The DE decomposition maintained acceptable BME detection specificity (⪅0.2 mL erroneously detected BME volume compared to 0.85 mL true BME volume) over +/-10% range of scatter magnitude errors, as long as the scatter shape was estimated without major distortions. Physical test bench experiments demonstrated successful discrimination of ~20% change in fat concentrations in trabecular bone-mimicking solutions of varying water and fat content.

Real-time imaging is highly desirable in image-guided radiotherapy, as it provides instantaneous knowledge of patient’s anatomy and motion during the treatment and enables online treatment adaptation to achieve the highest tumor targeting accuracy. Due to extremely limited acquisition time, only one or several X-ray projections can be acquired for real-time imaging, which poses a substantial challenge to localize the tumor from the scarce projections. For liver radiotherapy, such a challenge is further exacerbated by the diminished contrast between the tumor and the normal liver tissues. Here, we propose a framework combining graph neural network-based deep learning and biomechanical modeling to track liver tumor in real time from a single on-board X-ray projection. The liver tumor tracking is achieved in two steps. First, a deep learning network is developed to predict the liver surface deformation, using image features learned from the X-ray projection. Second, the intra-liver deformation is estimated through biomechanical modeling, using the liver surface deformation as the boundary condition to solve intra-liver tumor motion by finite element analysis. The accuracy of the proposed framework was evaluated using a dataset of 10 patients with liver cancer. The results show accurate liver surface registration from the graph-based neural network, which translates into accurate real-time, fiducial-less liver tumor localization (<1.3 mm localization error).

Today, the subcutaneous, minimally invasive procedures performed in interventional radiology are usually guided by 2D X-ray fluoroscopy. In 2D X-ray fluoroscopy a series of 2D X-ray images is displayed. For challenging procedures however, 3D X-ray fluoroscopy would be advantageous. In 3D X-ray fluoroscopy, a series of 3D images, which is reconstructed from a series of 2D X-ray images, is displayed. Because the number of images used for guiding an intervention is very high, little dose can be spent per 3D reconstruction of a 3D fluoroscopy. To save dose and to minimize motion artifacts, a reconstruction algorithm that requires very few X-ray projections is desirable. Earlier work showed that guidewires, stents and coils, which are commonly used in interventions, can be reconstructed using only four synthetic X-ray projections. The reconstruction from two or three X-ray projections was only studied briefly. In this work, we improve the method by using a more suitable neural network architecture and by using a multi-channel backprojection instead of a single-channel backprojection. We then apply the improved method to more realistic data measured in an anthropomorphic phantom. The results show that the method produces 3D reconstructions of stents and guidewires with submillimeter accuracy using only three measured X-ray projections.

Deformable motion is one of the main challenges to image quality in interventional cone beam CT (CBCT). Autofocus methods have been successfully applied for deformable motion compensation in CBCT, using multi-region joint optimization approaches that leverage the moderately smooth spatial variation motion of the deformable motion field with a local neighborhood. However, conventional autofocus metrics enforce images featuring sharp image-appearance, but do not guarantee the preservation of anatomical structures. Our previous work (DL-VIF) showed that deep convolutional neural networks (CNNs) can reproduce metrics of structural similarity (visual information fidelity - VIF), removing the need for a matched motion-free reference, and providing quantification of motion degradation and structural integrity. Application of DL-VIF within local neighborhoods is challenged by the large variability of local image content across a CBCT volume and requires global context information for successful evaluation of motion effects. In this work, we propose a novel deep autofocus metric, based on a context-aware, multi-resolution, deep CNN design. In addition to the inclusion of contextual information, the resulting metric generates a voxel-wise distribution of reference-free VIF values. The new metric, denoted CADL-VIF, was trained on simulated CBCT abdomen scans with deformable motion at random locations and with amplitude up to 30 mm. The CADL-VIF achieved good correlation with the ground truth VIF map across all test cases with R2 = 0.843 and slope = 0.941. When integrated into a multi-ROI deformable motion compensation method, CADL-VIF consistently reduced motion artifacts, yielding an average increase in SSIM of 0.129 in regions with severe motion and 0.113 in regions with mild motion. This work demonstrated the capability of CADL-VIF to recognize anatomical structures and penalize unrealistic images, which is a key step in developing reliable autofocus for complex deformable motion compensation in CBCT.

Cone-beam CT (CBCT) is widely used for guidance in interventional radiology but it is susceptible to motion artifacts. Motion in interventional CBCT features a complex combination of diverse sources including quasi-periodic, consistent motion patterns such as respiratory motion, and aperiodic, quasi-random, motion such as peristalsis. Recent developments in image-based motion compensation methods include approaches that combine autofocus techniques with deep learning models for extraction of image features pertinent to CBCT motion. Training of such deep autofocus models requires the generation of large amounts of realistic, motion-corrupted CBCT. Previous works on motion simulation were mostly focused on quasi-periodic motion patterns, and reliable simulation of complex combined motion with quasi-random components remains an unaddressed challenge. This work presents a framework aimed at synthesis of realistic motion trajectories for simulation of deformable motion in soft-tissue CBCT. The approach leveraged the capability of conditional generative adversarial network (GAN) models to learn the complex underlying motion present in unlabeled, motion-corrupted, CBCT volumes. The approach is designed for training with unpaired clinical CBCT in an unsupervised fashion. This work presents a first feasibility study, in which the model was trained with simulated data featuring known motion, providing a controlled scenario for validation of the proposed approach prior to extension to clinical data. Our proof-of-concept study illustrated the potential of the model to generate realistic, variable simulation of CBCT deformable motion fields, consistent with three trends underlying the designed training data: i) the synthetic motion induced only diffeomorphic deformations – with Jacobian Determinant larger than zero; ii) the synthetic motion showed median displacement of 0.5 mm in regions predominantly static in the training (e.g., the posterior aspect of the patient laying supine), compared to a median displacement of 3.8 mm in regions more prone to motion in the training; and iii) the synthetic motion exhibited predominant directionality consistent with the training set, resulting in larger motion in the superior-inferior direction (median and maximum amplitude of 4.58 mm and 20 mm, ⪆2x larger than the two remaining direction). Together, the proposed framework shows the feasibility for realistic motion simulation and synthesis of variable CBCT data.

Cardiac CT exams are some of the most complex CT exams due to need to carefully time the scan to capture the heart during the quiescent cardiac phase and when the contrast bolus is at its peak concentration. We are interested in developing a robust and autonomous cardiac CT protocol, using deep learning approaches to extract contrast timing and cardiac phase timing directly from pulsed projections. In this paper, we present a new approach to generate large amounts of clinically realistic virtual data for training deep learning networks. We propose a five-dimensional cardiac model generated from 4D cardiac coronary CT angiography (CTA) data for synthetic contrast bolus dynamics and patient ECG profiles. We apply deep learning to segment seven heart compartments and simulate intravenous contrast propagation through each compartment to insert contrast bolus. Additional augmentation techniques by randomizing a bolus curve, patient ECG profile, acquisition timing, and patient motion are applied to increase the amount of data that can be generated. We demonstrate good performance of the deep learning segmentation network, examples of simulated bolus curves using a realistic protocol, and good correspondence between virtually generated projections and real projections from patient scans.

Cardiac CT is a safe, accurate, non-invasive method widely employed for diagnosis of coronary artery disease (CAD) and planning therapeutic interventions. Even with state-of-the-art CT technology, calcium blooming artifacts may limit the accuracy of coronary stenosis assessment. A variety of solutions to reduce blooming artifacts have been proposed, including hardware improvements, protocol optimizations, and software deblooming techniques [1-6]. Hardware developments and clinical studies (for protocol optimization or training data generation) can be expensive, time-consuming, and impractical. Hence, there is an opportunity for a Virtual Clinical Trial (VCT) framework [7-8] to help researchers to evaluate the impact of various solutions on calcium blooming and to create training datasets for developing deep learning solutions for deblooming. In this paper, we present a new VCT framework for generating cardiac CT images with calcium blooming with a variety of CT hardware parameters, CT scan protocols and CT reconstruction kernels. As an example, we use the VCT framework to investigate the impact of three common scan and reconstruction parameters (X-ray tube voltage, focal spot size, and reconstruction kernel) on calcium blooming artifacts. We conclude that tube voltage and reconstruction kernel have the most direct impact on calcium blooming, which is consistent with earlier clinical reports [9-12].

Because of the ability to present molecular and functional information in organisms, nuclear medical imaging (NMI) is attracting more and more attention. Among NMI modalities, X-ray fluorescence computed tomography (XFCT) has the advantage that the tracers used in XFCT are not spontaneously decayed. The synthesis, storage of contrast agents is more convenient, the price of XFCT is much lower as well. However, XFCT usually has mechanical collimation to tell the incident photon direction, which results in the reduction of the detection efficiency. The Compton camera is an imaging modality, which does not need mechanical collimators in its structure, which makes Compton cameras have high detection efficiency. Therefore, it is a great idea to use Compton camera-based imaging systems to realize X-ray fluorescence (XF) imaging. In this work, the first XFCC imaging system in the laboratory environment is established, which consists of a 150keV X-ray tube and a single-layer Compton camera system based on the Timepix3 photon-counting detector (PCD). The element Gd (43keV) is used as the XF element. The first imaging reconstruction results of the XFCC system are represented.

Breast cancer is the most common malignancy in women. Unfortunately, even though screening programs have helped to increase survival rates, the number of false positives and false negatives remains high. Phase-contrast X-ray CT is a promising imaging technique which could improve breast cancer diagnosis by combining the high three-dimensional resolution of conventional CT with higher soft-tissue contrast. Grating Interferometry CT (GI-CT) arguably has the highest chance to make the transition to clinical practice. Unfortunately though, obtaining high-quality images is challenging. Grating fabrication defects and photon starvation lead to high noise amplitudes in the measured data. Moreover, the highly ill-conditioned differential nature of the GI-CT forward operator renders the inversion from corrupted data even more cumbersome. In this article we report on a novel regularized iterative reconstruction algorithm with a powerful data-driven regularization strategy to tackle this challenging inverse problem. In particular, we present an algorithm that combines the L-BFGS optimization scheme with a Plug-and-Play denoiser parameterized by a deep neural network and empirically show that the proposed method achieves high quality images, both on simulated data as well as on real measurements.

Scatter artifacts is one of the limitations of cone-beam computed tomography (CBCT) image quality, this paper proposed a novel scatter correction method for CBCT via combining the deep learning and forward projection algorithm. This method can be mainly divided follow steps: Firstly, raw projections were used to reconstruct raw CBCT via FDK algorithm, and then the raw CBCT was processed by CycleGAN network to generate synthetic CT, after that the synthetic CT was used to forward projection based on the Beer’s laws to generate scatter free primary projections. The raw scatters can be estimated via subtracting the primary projections from the raw projections, and then a median and low-pass Gaussian filter was used to smooth the raw scatters. Finally, The scatter corrected projections can be acquired by subtracting the filtered scatters form raw projections. The study results of pelvis and chest validated the effective of proposed correction method in reducing scatter artifacts of CBCT. Compared with uncorrected CBCT and CBCT corrected by scatter kernel superposition(SKS) method, the proposed method can more effectively improve the quality of reconstructed CBCT, reducing the CT number errors and increasing contrast-to-noise (CNR) and so on. All results show that this method has strong scatter artifacts restriction ability, so it has significant promising for clinical application.

We have developed VSHARP®, a suite of scatter correction solutions that have been incorporated into the commercially available cone-beam software development toolkit, CST (Varex Imaging, Salt Lake City, UT) enabling scatter correction to be applied as part of an entire CBCT reconstruction pipeline. The suite includes 2D VSHARP®, a deconvolution correction using asymmetric Gaussian kernels, 2D VSHARP-ML, a U-NET machine-learning correction, and 3D VSHARP®, a correction using a rapid finite-element Linear Boltzmann Transport Equation (LBTE) solver to estimate scatter in a manner similar to traditional stochastic Monte Carlo (MC) simulations. Of the three corrections, 3D VSHARP is the most accurate and flexible since it can be readily applied to arbitrary scanner geometries, protocols, and scan parts while the 2D VSHARP models may need to be regenerated for each configuration. On the other hand, 3D VSHARP is inherently slower since a minimum of two reconstruction passes are needed and the LBTE solver, while much faster than traditional MC, is still computationally intensive. The goal of this work was to minimize LBTE run times for (typically large) industrial datasets by optimizing parameter settings, particularly the choice of the sampling grid dimensions. This was achieved by applying a multi-objective genetic algorithm to find the Pareto front characterizing the tradeoff between speed and accuracy and identifying key operating points on the curve. Testing with 720 frames of 3720x3720 projection data to make a reconstruction volume of size 500x500x600, we found that excellent image quality can be obtained by using a coarse scatter grid size of 27x27x32 volume and 44x44 detector and a primary grid size of 246x246 x295 volume and 295x295 detector, both over 42 frames for a grand total of 21 seconds LBTE computation time. We show the Pareto characterization, as well as demonstrations of 3D VSHARP image quality with significantly reduced scatter-induced artifacts such as streaking and shading.

NeuralCT [1] has been recently proposed as an implicit neural representation-based image reconstruction that can produce time-resolved images from CT sinograms and reduce motion artifacts, even when undergoing complex motions. NeuralCT does not require the prior motion model or estimation of object motion. Instead, it utilizes a network to implicitly represent the time-varying object boundary by singed distance function and optimizes the network via differentiable rendering. In this work, we modify the NeuralCT framework to reconstruct scenes that have multiple moving objects with distinct attenuation levels. We show that the performance of NeuralCT reconstruction depends on the quality of the initialization of the network (in this case, object segmentation in motion-corrupted FBP image). We show how spatially aware object segmentation can improve motion-corrected reconstruction in moving objects with multiple attenuation levels despite high angular motion and complex topological changes.

Photon-counting detector (PCD) CT promises to improve routine CT imaging applications with higher spatial resolution, lower levels of noise at fixed dose, and improved image contrast while providing spectral information with every scan. We propose and demonstrate a novel application of PCD imaging in a preclinical model of head and neck squamous cell carcinoma: spectral perfusion imaging of cancer. To handle the high dimensionality of our data set (3D volumes at 12 perfusion time points times 4 energy thresholds), we update our previously proposed multi-channel iterative reconstruction algorithm to handle the perfusion reconstruction problem, and we propose an extension which adds patch-based singular value thresholding (pSVT) along the perfusion dimension. Adding pSVT reduces noise by an additional 45% relative to our standard algorithm, which itself reduces noise by 2-7 times relative to analytical reconstruction. Preliminary analysis suggests that the addition of pSVT does not negatively impact material decomposition accuracy or image spatial resolution. Notable weaknesses of this preliminary study include relatively high contrast agent dose (0.5 mL ISOVUE-370 over 10 seconds), ionizing radiation dose (~570 mGy), and computation time (2.9 hours, no pSVT; 11 hours with pSVT); however, following from our past work, our reconstruction algorithm may be an ideal source of training labels for supervised deep learning applied to computationally cheap analytical reconstructions.

Dynamic tomography finds its usage in certain important applications. The reconstruction problem could be cast in the Bayes framework as a time series analysis problem, such that the Kalman filter (KF) could come into play. A series of drawback of such an approach lies in its high computational and storage complexity. Dimension reduction Kalman filter (DR-KF) method has been proposed in the literature to relieve such a pain. However, our tests show that DR-KF results in heavy ringing artifacts. To solve this dilemma, in this paper, we propose to approximate the regularization term involving the precision matrix with a spatial regularization term plus a temporal regularization term, such that the space and time complexity are greatly reduced. Besides, to get the original Kalman filter into play, we propose a blocked KF for backward smoothing, i.e. split the reconstructed image slices into small overlapping blocks, and the KF is applied on each block. The blocked KF has much smaller computational and storage complexity, and fits well for parallel computations. Numerical experiments show that the proposed approach achieves better reconstructions while calling for much smaller computational resources.

Deep learning (DL) based methods have been widely adopted in computed tomography (CT) field. And they also show a great potential in photon-counting CT (PCCT) imaging field. They usually require a large quantity of paired data to train networks. However, it is time-consuming and expensive to collect such large-scale PCCT dataset. In addition, lots of energy-integrating detector (EID) data are not yet included in the DL-based PCCT reconstruction network training. In this work, to address the issue of limited PCCT data and take advantage of labeled EID data, we propose a novel unsupervised full-spectrum-knowledge-aware DL-based network (FSANet), which contains supervised and unsupervised networks, to produce high-quality PCCT images. Specifically, the supervised network is trained based on paired EID dataset and serves as the prior knowledge to regularize the unsupervised PCCT network training. Moreover, a data-fidelity term for characterizing the PCCT image characteristics is constructed as a self-supervised term. Finally, we train the PCCT network with the prior knowledge and self-supervised terms following an unsupervised learning strategy. Numerical studies on synthesized clinical data are conducted to validate and evaluate the performance of the presented FSANet method, qualitatively and quantitatively. The experimental results demonstrate that presented FSANet method significantly improves the PCCT image quality in the case of limited photon counts.

Computed tomography (CT) is a non-invasive means of localizing a region of interest within an object, which enables the investigation of soil distributions and localized flow processes within soil pore systems. CT scanning allows for cross-sectional successions that provide visibility within the environment of plant samples. Knowledge of the characteristics of the soil pore system is essential for evaluating various processes that take place between root-soil interactions. In this study, we investigate the potential application of a high-resolution amorphous selenium (a-Se) direct conversion detector on complementary metal-oxide-semiconductor (CMOS) readouts for micro-CT scanning of a soil matrix to image the status of aggregation and networks of pore spaces within intact soil. The combination of the intrinsic high spatial resolution of a-Se and small pixel CMOS readouts provide detailed visualization in the soil aggregates of the plant samples. The x-ray energy and plant soil thickness were varied during the investigation for imaging the root-soil. A 10 μm spatial resolution and noise limited performance of eight photons/pixel at 20 keV were achieved. High attenuation of x-rays in thick soil poses challenges however fine details are observable in thinner samples and care should be taken when choosing soil thickness and container material.

Multi-energy CT conducted by photon-counting detectors has a wide range of applications, especially in multiple contrast agent imaging. However, multi-energy CT imaging suffers from higher statistical noise because of increased energy bin numbers. Our team has proposed the dynamic dual-energy CT imaging mode and the corresponding iterative imaging algorithms to solve this problem. The multi-energy projections and reconstructions calculated from the dynamic dual-energy CT data are less noisy than the static multi-energy CT, which has been verified by sufficient numerical simulations and experiments. However, a rigorous mathematical derivation has not been conducted to explain why dynamic dual-energy CT is better than static multi-energy CT in reducing statistical noise. In this work, we drive the noise model of the dynamic dual-energy CT to explain the reason. The reason is: compared to the multi-energy projections that are directly measured from a static multi-energy CT, the multi-energy projections, which are calculated from the dynamic dual-energy CT data, have the same expectation, but the variance is lower.

In recent years, cone-beam X-ray luminescence computed tomography (CB-XLCT) has drawn much attention with the development of X-ray excited nanophosphors. Compared with traditional bio-optical imaging modalities such as bioluminescence tomography (BLT) and fluorescence molecular tomography (FMT), CB-XLCT can effectively improve imaging sensitivity and depth because of the reduction of background fluorescence and the high penetrability of X-rays. However, due to high degree of scattering of light through biological tissues, the reconstruction of CB-XLCT is inherently ill-conditioned. To solve the ill-posed inverse problem, appropriate priors or regularizations are needed to facilitate the reconstruction. Based on the fact that adjacent pixels generally have the same or similar concentration and in order to further balance the degree of regional smoothness and edge sharpening, a prior information model based on Huber Markov Random Field (HuMRF) was established to constrain the reconstruction process of CB-XLCT. Mice experiments indicate that compared with the traditional ART and ADAPTIK method, the proposed method could improve the image quality of CB-XLCT significantly in terms of target shape, localization accuracy and image contrast.

In a cone-beam CT system, the use of bowtie-filter may induce artifacts in the reconstructed images. Through a Monte-Carlo simulation study, we confirm that the bowtie filter causes spatially biased beam energy difference thereby creating beam-hardening artifacts. We also note that cupping artifacts in conjunction with the object scatter and additional beam-hardening may manifest. In this study, we propose a dual-domain network for reducing the bowtie-filter induced artifacts by addressing the origin of artifacts. In the projection domain, the network compensates for the filter induced beam-hardening effects. In the image domain, the network reduces the cupping artifacts that generally appear in cone-beam CT images. Also, transfer learning scheme was adopted in the projection domain network to reduce the total training costs and to increase utility in the practical cases while maintaining the robustness of the dual-domain network. Thus, the pre-trained projection domain network using simple elliptical cylinder phantoms was utilized. As a result, the proposed network shows denoised and enhanced soft-tissue contrast images with much reduced image artifacts. For comparison, a single image domain U-net was also implemented as an ablation study. The proposed dual-domain network outperforms, in terms of soft-tissue contrast and residual artifacts, a single domain network that does not physically consider the cause of artifacts.

An important goal in modern CT imaging is reducing the dose delivered to patients especially to risk-relevant organs. This work compares the clinically applied dose reduction techniques mAs-minimizing tube current modulation (mAsTCM) and a typical organ-specific TCM (osTCM, here: X-Care, Siemens Healthcare) with a novel radiation risk-minimizing tube current modulation (riskTCM) with a focus on the dose delivered to the female breast. The mAsTCM minimizes the mAs product as a surrogate parameter for the patient dose but does not consider the different organs’ risks. In contrast, osTCM aims to minimize the dose delivered to the female breast by reducing the tube current for anterior projections. The riskTCM minimizes the patient risk by minimizing the effective dose to the patient, which is done by taking accounting for the organ doses. In this study, the dose reduction effect of the TCM techniques is compared by simulations based on clinical CT Scans. Results: riskTCM reduces the effective dose by up to 35% in comparison to mAsTCM, and by up to 30% in comparison to osTCM depending on the anatomical region and the patient.

Image mean and covariance required for a model observer are usually calculated by the statistical method using image samples, which is hard to acquire in reality. Although some analytical methods are proposed to estimate image covariance from a single projection, these methods are of high computational cost for large-dimensional images (e.g., 512×512), and images of large dimension are commonly required. Considering the covariance used for a model observer is the covariance of the channel response, whose dimension is much smaller than the image covariance, we aim to obtain the covariance of small-dimensional channel response directly from its projection. Channel filters are applied to the analytical projection to image (Prj2Img) covariance estimation method to derive the analytical projection to channel response (Prj2CR) covariance estimation method, which successfully reduces the computational cost and connects the covariance of projection and channel response. In addition, a transition matrix is introduced in Prj2CR method to stabilize the connections. The transition matrix mainly depends on channel filters, not the system, phantom, and reconstruction algorithm, which means it can be calibrated by small-dimensional reconstructions and then applied to any situation with a same channel filter. We validate the feasibility and utility of the proposed Prj2CR method by simulations. 128×128 reconstructions from qGGMRF-WLS are adopted for calibration, while 512×512 reconstructions are used for validation. SNR of CHO is chosen as the figure of merit for performance evaluation, and the covariance estimated by 290 image samples are used as the reference. Results show that the SNR by the Prj2CR method is within 95% confidence interval of the SNR* by 290 image samples, indicating that the proposed method accords with statistical method. The Prj2CR method may be beneficial for subjective image quality assessment since it only needs a single sample of projection and has low computational cost.

Although photon counting detectors (PCD) can offer numerous benefits for CT imaging, it is difficult to generate accurate material decompositions from photon counting (PC) CT images due to spectral distortions. In this work, we present a deep learning (DL) approach for material decomposition from PCCT. To produce training and testing data for this study, we scanned two ex-vivo mice using a PCD scan protocol with a dose of 36 mGy and a multi-EID scan protocol with a dose of 296 mGy. PCD images were reconstructed using filtered backprojection. EID images were reconstructed using an iterative algorithm to reduce noise, and decomposed into iodine (I), Compton scattering (CS), and photoelectric effect (PE) material maps by a matrix inversion approach. We then trained a convolutional neural network with a 3D U-net structure using PCD images as inputs and multi-EID material maps as labels, and evaluated its performance. The 3D U-net predictions provided substantially lower RMSE compared to decomposition from PCD images using a matrix inversion approach. Measurements from iodine vials in the test set showed that 3D U-net predictions gave mean values within 0.6 mg/mL of the mean values from the multi-EID material maps and much lower standard deviation than PCD material map measurements. Our results show that the trained 3D U-net enables low-noise, quantitatively accurate material decomposition from a low dose PCD scan.

Based on the X-ray physics in computed tomography (CT) imaging, the linear attenuation coefficient (LAC) of each human tissue is described as a function of the X-ray photon energy. Different tissue types (i.e. muscle, fat, bone, and lung tissue) have their energy responses and bring more tissue contrast distribution information along the energy axis, which we call tissue-energy response (TER). In this study, we propose to use TER to generate virtual monoenergetic images (VMIs) from conventional CT for computer-aided diagnosis (CADx) of lesions. Specifically, for a conventional CT image, each tissue fraction can be identified by the TER curve at the effective energy of the setting tube voltage. Based on this, a series of VMIs can be generated by the tissue fractions multiplying the corresponding TER. Moreover, a machine learning (ML) model is developed to exploit the energy-enhanced tissue material features for differentiating malignant from benign lesions, which is based on the data-driven deep learning (DL)-CNN method. Experimental results demonstrated that the DL-CADx models with the proposed method can achieve better classification performance than the conventional CT-based CADx method from three sets of pathologically proven lesion datasets.

Successful image reconstruction in computed tomography (CT) relies on the completeness of the projections. If the patient does not fit in the field of measurement, the projections are truncated causing cupping artifacts in the image and a diminished field of view (FOV). In order to restore the CT values and extend the FOV, the projections have to be completed, for example via an extrapolation. The discrete algebraic reconstruction technique (DART) has shown its efficacy in reconstructing discrete images from insufficient raw data. In this work, we use DART images as a prior for projection completion of clinical CT scans. We compare our method to the conventional adaptive detruncation (ADT) and evaluate the RMSE inside and outside the FOV along with the Dice score.

We present a feasibility study on extracting the respiratory surrogate signal (RSS) from cone-beam computed tomography (CBCT) projections using a supervised convolutional neural network (CNN) model. Determining the intrinsic RSS instead of using an external surrogate signal, provided by optical tracing hardware such as the Real-time Position Management (RPM) system, has the advantage that it spares patient setup time and hence permits faster 4D CBCT acquisition. Another convenience of such an approach is that it can be applied retrospectively without special preparation or equipment. For the implementation, we made use of the MONAI open source library. Our model is based on a modified version of the MONAI regressor class. We trained the model using CBCT projections with the corresponding RSS as recorded by an external marker block using the RPM system. The model is to deduce the RSS given the CBCT. Using a specific dataset with CBCT from anesthetized animals breathing with the help of a mechanical ventilator, results show a good correlation between the actual and predicted RSS. Unlike the Amsterdam Shroud algorithm, our method shows promising results to predict the normalized amplitude of the breathing signal. Further work could extend the model to permit RSS prediction for its online use in radiation therapy or detection of sudden motion deteriorating CBCT image quality. To conclude, we have made a first step towards proving the concept which consists in using a deep learning model to extract the RSS out of acquired CBCT projection images. The approach is promising but requires more work for robustness, i.e. for sufficient accuracy both in the frequency and normalized amplitude extraction.

Wide-coverage detector CT and ultra-high-resolution (UHR) detector CT are two important features for current cardiac imaging modalities. The former one enables the scanner to finish a whole heart image scan in one bed position; the latter one gives superior resolution in fine structures such as stenoses, calcifications, implemented stents, and small vessel boundaries. However, no commercially available scanner has both these features in one scanner as of today. Herein, we propose to use existing UHR-CT data to train a super resolution (SR) neural network and apply the network in a wide-coverage detector CT system. The purpose of the network is to enhance the system resolution performance and reduce the noise while maintaining its wide-coverage feature without additional hardware changes. Thirteen UHR-CT patient datasets and their simulated-normal-resolution pairs were used for training a 3D residual-block U-Net. The modulation transfer function (MTF) measured from Catphan phantom scans showed the proposed super-resolution aided deep learning-based reconstruction (SR-DLR) improved the MTF resolution by relative ~30% and ~10% as compared to filtered-back projection and model-based iterative reconstruction approaches. In real patient cases, the SR-DLR images show better noise texture and enhanced spatial resolution along with better aortic valve, stent, calcification, and soft tissue features as compared to other reconstruction approaches.

Dual-energy CT (DECT) of limited-angular ranges (LARs) collects data from angular ranges smaller than π for low- and high-kVp scans, and thus may potentially be exploited for reducing scanning time and radiation dose and for avoiding collision between the imaged object and the moving gantry of the scanner. Image artifacts resulting from beam hardening (BH) and limited-angular range (LAR) can be suppressed by using the data-domain decomposition and the directional-total-variation (DTV) algorithm for image reconstruction. In this work, we investigate two-orthogonal-arc (TOA) scanning configuration with overlapping arcs for collecting LAR DECT data, in an effort to reduce LAR artifacts and improve quantitative accuracy of estimated physical quantities. The TOA configuration consists of two arcs, of equal LAR, whose centers are positioned 90° apart, and is designed to reduce the ill-conditionedness of the imaging system matrix. The data are decomposed into basis sinograms, from which basis images are reconstructed using the DTV algorithm. Visual inspection of the monochromatic images and quantitative estimation of the effective atomic numbers suggest that the TOA configuration, as compared to the single-arc (SA) configuration of the same total angular range, can help reduce remaining LAR artifacts and bias in the estimated atomic number relative to the reference values from the full-angular-range data of 360° .

A novel automated workflow for the recovery of image resolution using deep convolutional neural networks (CNNs) trained using spatially registered multiscale data is presented. Spatial priors, coupled with high order voxel-based image registration, are used to correct for uncertainties in image magnification and position. A network is then trained to remove the effects of point spread from the low-resolution data, improving resolution while reducing image noise & artefact levels. While benchmarking on real materials, including biological, materials science and electronics samples, we find that resolution recovery improves quantitative and qualitative measurements, even if certain image details cannot be easily identified from the original low-resolution data.

Flat panel detector (FPD) based cone-beam computed tomography (CT) has made tremendous progress these days, with many new medical and industrial applications keeping emerging from diagnostic imaging and image guidance for radiotherapy and interventional surgery. However, current cone-beam CT (CBCT) is still suboptimal for head scan whose requirement for image quality is extremely strict. Recently, the dual-layer flat panel detector technology is under development and is promising to further advance CBCT from qualitative anatomic imaging to quantitative dual-energy CT. Its potential of enabling head CBCT applications has yet been investigated. The relatively moderate energy separation from the dual-layer FPD and the overall low signal level especially for the bottom layer detector, raise significant challenges in performing high quality dual-energy material decomposition. In this work, we propose a physics based material decomposition algorithm that attempts to fully use the detected X-ray signals and prior-knowledge behind head CBCT using dual-layer FPD. Specifically, projection data from the two layers of detector are first adaptively combined to generate conventional CT images with reduced noise. A physics model based dual-layer multi-material spectral correction (dMMSC) is then developed to make the combined image reconstruction beam-hardening free. After a regular projection-domain material decomposition (MD) being conducted, the corresponding beam-hardening free projections from the dMMSC will be taken as a guidance to further enhance the dual-layer MD performance, leading to significantly improved robustness of MD and suppressed low-signal artifact in our preliminary results.

Traditional model-based image reconstruction (MBIR) methods combine forward and noise models with simple object priors. Recent application of deep learning methods for image reconstruction provides a successful data-driven approach to addressing the challenges when reconstructing images with measurement undersampling or various types of noise. In this work, we propose a hybrid supervised-unsupervised learning framework for X-ray computed tomography (CT) image reconstruction. The proposed learning formulation leverages both sparsity or unsupervised learning-based priors and neural network reconstructors to simulate a fixed-point iteration process. Each proposed trained block consists of a deterministic MBIR solver and a neural network. The information flows in parallel through these two reconstructors and is then optimally combined, and multiple such blocks are cascaded to form a reconstruction pipeline. We demonstrate the efficacy of this learned hybrid model for low-dose CT image reconstruction with limited training data, where we use the NIH AAPM Mayo Clinic Low Dose CT Grand Challenge dataset for training and testing. In our experiments, we study combinations of supervised deep network reconstructors and sparse representations-based (unsupervised) learned or analytical priors. Our results demonstrate the promising performance of the proposed framework compared to recent reconstruction methods.

X-ray imaging is the most popular medical imaging technology. While X-ray radiography is rather cost-effective, tissue structures are superimposed along the X-ray paths. On the other hand, computed tomography (CT) reconstructs internal structures but CT increases radiation dose, is complicated and expensive. Here we propose ”X-ray dissectography” to extract a target organ digitally from few radiographic projections for stereographic and tomographic analysis in the deep learning framework. As an exemplary embodiment, we propose a general X-ray dissectography network, a dedicated X-ray stereotography network, and the X-ray imaging systems to implement these functionalities. Our experiments show that X-ray stereography can be achieved of an isolated organ such as the lungs in this case, suggesting the feasibility of transforming conventional radiographic reading to the stereographic examination of the isolated organ, which potentially allows higher sensitivity and specificity, and even tomographic visualization of the target. With further improvements, X-ray dissectography promises to be a new X-ray imaging modality for CT-grade diagnosis at radiation dose and system cost comparable to that of radiographic or tomosynthetic imaging.

Purpose: to investigate image quality of the ultra-high-resolution (UHR) mode of a dual-source photon-counting CT scanner in visualizing mixed (soft and hard) coronary artery plaques. Materials and methods: We scanned a custom-made phantom with 10 mixed plaques of various sizes and compositions. Each scan was repeated three times. Images were reconstructed with FBP, and model-based quantum iterative reconstruction (QIR). Image quality was investigated by measuring mean CT numbers, noise standard deviation (SD), and by line profiles. Results: UHR mode provided sharper difference between soft and hard plaques, and the lumen by reducing blooming artifacts. Furthermore, it improved the true CT number of the values by reducing partial volume However, SD of noise increases by a factor of ~8 in FBP reconstructions at thinnest slice thickness (0.2 mm). Quantum iterative reconstruction algorithm reduced image noise x4 of the SR FBP without any apparent loss of spatial resolution. Conclusion: UHR PCCT improves plaque characterization through improved spatial resolution which results in lower blooming artifacts and partial volume effects. The increase in image noise can be mitigated by using model-based iterative reconstruction algorithms without any loss of spatial resolution. Depending on the imaging task, further noise reduction can be achieved by reconstructing thicker slices. A more detailed investigation with noise power spectrum analysis and observer model studies is warranted.

The x-ray computed tomography (CT) images with low dose are noisy and may contain photon starvation artifacts. The artifacts are location and direction dependent. Therefore, the common shift-invariant denoising filters do not work well. The state-of-the-art methods to process the low-dose CT images are image reconstruction based; they require the raw projection data. In many situations, the raw CT projections are not accessible. This paper suggests a method to denoise the low-dose CT image using the pseudo projections generated by the application of a forward projector on the low-dose CT image. The feasibility of the proposed method is demonstrated by real clinical data.

Metal objects in the field of view cause artifacts in the image, which manifest as dark and bright streaks and degrade the diagnostic value of the image. Standard approaches for metal artifact reduction are often unable to correct these artifacts sufficiently or introduce new artifacts. We propose a new data-based method to reduce metal artifacts in CT images applying conditional Generative Adversarial Networks to the corrupted data. A generator network is applied directly to the corrupted projections by the metal objects to learn the corrected sinogram data. Further, two discriminator networks are used to evaluate the image quality of the enhanced data from the generator. The method was initially developed based on a supervised approach. However, there is usually no ground truth for actual clinical data without artifacts, which is needed to train the networks. Therefore, the method was further improved to train an unsupervised network, i.e., without the ground truth. In addition the input data, the neighboring slices and the stochastic components of the image are included using the latent space representation of the data. The results show that the trained generator network can reasonably replace the missing projection data and reduce the artifacts in the reconstructed image.

Artifact correction is a great challenge in cardiac imaging. During the correction of coronary tissue with motion-induced artifacts, the spatial distribution of CT value not only shifts according to the motion vector field (MVF), but also shifts according to the volume change rate of the local voxels. However, the traditional interpolation method does not conserve the CT value during motion compensation. A new sample interpolation algorithm is developed based on the constraint of conservation of CT value before and after image deformation. This algorithm is modified on the existing interpolation algorithms and can be embedded into neural networks with deterministic back propagation. Comparative experimental results illustrate that the method can not only correct motion-induced artifacts, but also ensure the conservation of CT value in the region of interest (ROI) area, so as to obtain corrected images with clinically recognized CT value. Both effectiveness and efficiency are proved in forward motion correction process and backward training steps in deep learning. Simultaneously, the visualized motion vector field transparentizes the correction process, making this method more interpretable than the existing image-based end-to-end deep learning method.

An optimization-based image reconstruction framework is developed specifically for bone imaging. This framework exploits voxel-sparsity by use of ℓ₁-norm image regularization and it enables image reconstruction from sparse-view cone-beam computed tomography (CBCT) acquisition. The effectiveness of the voxel-sparsity regularization is enhanced by using a blurred image representation. Ramp-filtering is included in the data discrepancy term and it has the effect of acting as a preconditioner, reducing the necessary number of iterations. The bone image reconstruction framework is demonstrated on CBCT data taken from an equine metacarpal condyle specimen.

Cardiac CT exams are some of the most complex CT exams due to the need to carefully time the scan to capture the heart during a quiescent cardiac phase and when the intravenous contrast bolus is at its peak concentration in the left and/or right heart. We are interested in developing a robust and autonomous cardiac CT exam, using deep learning approaches to extract contrast and cardiac phase timing directly from projections. In this paper, we present a new approach to estimate contrast bolus timing directly from a sparse set of CT projections. We present a deep learning approach to estimate contrast agent concentration in left and right sides of the heart directly from a set of projections. We use a virtual imaging framework to generate training and test data, derived from real patient datasets. We finally combine this with a simple analytical approach to decide on the start of the cardiac CT exam.

The perfusion imaging using C-arm CT could be used intraoperatively for liver cancer treatment planning and evaluation. To deal with undersampled data due to slow C-arm CT rotation and pause between the rotations, we applied model-based reconstruction methods. Recent works using the time separation technique with an analytical basis function set have led to a significant improvement in the quality of C-arm CT perfusion maps. In this work we apply the time separation technique with a prior knowledge basis function set extracted using singular value decomposition from CT perfusion reconstructions. On C-arm CT liver perfusion scan simulated based on the real CT liver perfusion scan we show that the bases extracted from only two CT perfusion scans are capable of modeling the C-arm CT data correctly.

Low-dose CT reconstructions suffer from severe noise and artifacts. Many methods have been proposed to increase the ratio between image quality and radiation dose through incorporating various priors. By learning priors in labelled data-set, neural network methods have achieved great success for this purpose. In CT applications, however, paired training data-sets are rarely available or difficult to obtain. Recently, unsupervised learning has attracted a lot of attention. Along this line, Noise2Inverse, an unsupervised neural network architecture, has shown the possibility of applying unsupervised learning for low-dose CT reconstructions. When the training data-sets (unlabelled) are not large enough or the training is insufficient, however, Noise2Inverse might perform not well. Another important issue is that network methods might suffer from intrinsic instability. In this regard, we propose to hybrid neural networks, especially the Noise2Inverse architecture, with traditional optimization models such that hand-crafted priors come into play as a remedy. Numerical experiments show that the proposed architecture improves Noise2Inverse in terms of both quality measures PSNR and SSIM, especially in the case of inadequate training.

Cerebral perfusion computed tomography (CPCT) imaging provides a rapid and accurate noninvasive measurements of the acute stroke by generating hemodynamic parameter maps with a qualitative and quantitative way. However, due to it performs a multiple consecutive scanning protocol at one area of the head, the radiation exposure is relatively higher than a routine protocol. And lowering radiation dose in CPCT protocol would increase the amount of noise and hence influence hemodynamic parameters for patients with acute stroke. Some advanced methods have been proposed and show a great potential in noise suppression for low-dose CPCT imaging. And most of them assume that the embedded noise obeys an independent and identically distribution (i.i.d), but the noise may be more complicated in practical scenarios. In this work, we first analyze the noise properties in low-dose CPCT images. And then present a novel perfusion deconvolution method with a self-relative structure similarity information and a mixture of Gaussians (MoG) noise model (named SR-MoG) to accurately estimate the hemodynamic parameters directly at the low radiation exposure. Experiments implemented on digital brain perfusion phantom verify that the presented SR-MoG method can achieve promising gains over the existing deconvolution approaches.

Deep learning-based image denoising and reconstruction methods have shown promising results for low-dose CT. When high-quality reference images are not available for training the network, researchers found a powerful and effective counterpart called Noise2Noise, which trains the neural network using paired data with independent noise. However, it is uncommon to have paired CT scans with independent noise (e.g., from two scans). In this paper, a method is proposed to generate such paired data for potential usage in deep learning training by simultaneously simulating a low-dose image at arbitrary dose level and an image with independent noise from a single CT scan. Their independence is investigated both analytically and numerically. In our numerical study, a Shepp-Logan phantom was utilized in MATLAB to generate the ground-truth, normal-dose, and low-dose images for reference. Noise images were obtained for analysis by subtracting the ground-truth from the noisy images, including the normal-dose/low-dose images and the paired products of our proposed method. Our numerical study matches the analytical results very well, showing that the paired images are not correlated. Under an additional assumption that they form a bivariate normal distribution, they are also independent. The proposed method can produce a series of paired images at arbitrary dose level given one CT scan, which provides a powerful new method to enrich the diversity of low-dose data for deep learning.

Grating-based phase-contrast and dark-field X-ray imaging is a promising technology for improving the diagnosis and imaging capabilities of breast cancer and lung diseases. While traditional X-ray techniques only consider the attenuation coefficient, phase-contrast and dark-field imaging are also capable of measuring the refractive index decrement and the so-called linear diffusion coefficient, a measure of a sample’s small-angle scattering strength. Consequently, the technique provides additional information about the micro-structure of a sample. While it is already possible to perform human chest dark-field radiography, it is assumed that its diagnostic value increases when performed in a tomographic setup. The thereby acquired three-dimensional mappings of the three modalities yield detailed information about morphological changes without being obscured by overlaying structures. This work presents the sample data processing and reconstruction pipeline of the first human-sized clinical dark-field CT system. In this novel setting we require a processing concept which is (1) compatible with continuous rotation, (2) can compensate for perturbances induced by system vibrations, and (3) still enables short processing and reconstruction times. An advanced sliding window approach was chosen for the sample data extraction to meet requirements (1) and (3). Furthermore, we present the corrective measures that have to be applied in the employed processing and reconstruction algorithms to mitigate the effects of vibrations and deformations of the interferometer gratings. The developed techniques are shown to successfully reduce the emergence of artefacts in the reconstructed images.

Photon counting spectral CT (PCCT) can produce reconstructed attenuation maps in different energy channels, reflecting energy properties of the scanned object. Due to the limited photon numbers and the non-ideal detector response of each energy channel, the reconstructed images usually contain much noise. With the development of Deep Learning (DL) technique, different kinds of DL-based models have been proposed for noise reduction. However, most of the models require clean data set as the training labels, which are not always available in medical imaging field. Inspiring by the similarities of each channel's reconstructed image, we proposed a self-supervised learning based PCCT image enhancement framework via multi-spectral channels (S 2MS). In S 2MS framework, both the input and output labels are noisy images. Specifically, one single channel image was used as output while images of other single channels and channel-sum image were used as input to train the network, which can fully use the spectral data information without extra cost. The simulation results based on the AAPM Low-dose CT Challenge database showed that the proposed S²MS model can suppress the noise and preserve details more effectively in comparison with the traditional DL models, which has potential to improve the image quality of PCCT in clinical applications.

Often, the artifacts caused by high-density objects degrade the quality of the image with streaks and information loss in CT imaging. In recent years, machine learning has proven itself a powerful tool to resolve some of the challenges faced in reducing metal artifacts. In this work, a novel method of metal artifact reduction (MAR) without metal segmentation by using a CNN network is proposed. The approach focuses on removing the need for the sensitive metal segmentation step to improve robustness and aims to tackle beam hardening directly in the sinogram domain. In the proposed method, we trained the network with sinogram pairs that include metal objects and those that include virtual non-metal (VNM) replacement objects. A VNM object is designed to be less dense than metal but more dense than soft tissue. The novelty of this method lies in the sinogram-to-sinogram training without the need for metal segmentation by replacing the metal object to a virtual non-metal object in the sinogram to reduce beam hardening and successfully compensate for the information loss.

Deep learning has achieved great success in many medical imaging tasks without explicit solutions. In this work, learning method was applied to dual-energy cone-beam CT imaging. We proposed a Residual W-shape Network (ResWnet). ResWnet consists of three modules: scatter correction module 𝒮, material decomposition module ℳ, decomposition denoising module 𝒟 . Both 𝒮 and 𝒟 use ResWnet architecture, and this lightweight model fuses multi-level features, achieving satisfied performance with a small number of parameters. 𝒮 acts on dual-energy attenuation projections to reduce the scatter contaminations, and 𝒟 acts on material composition projections to suppress the noise. ℳ links the modules 𝒮 and 𝒟, and is used for domain transform from attenuation projections to material projections. This process could be approximated by polynomials with pre-calibrated parameters, that is, ℳ is a known operator in proposed network with no trainable parameters. This helps to reduce model parameters and improve the performance with small training dataset. Using public head CT dataset, we simulated dual-energy cone-beam CT projections and material projections. Proposed ResWnet was trained, validated and tested on this simulated dataset, verifying its effectiveness in projection-domain scatter correction and low-noise decomposition.

X-ray computed tomography (CT) is an invaluable imaging technique for non-invasive medical diagnosis. However, for soft tissue in the human body the inherent small difference in attenuation limits its significance. Grating-based X-ray phase-contrast is a relatively novel imaging method which detects additional interaction mechanisms between photons and matter, namely refraction and small-angle scattering, to generate additional images with different contrast. The experimental setup involves a Talbot-Lau interferometer whose susceptibility to mechanical vibrations hindered acquisition schemes suitable for clinical routine in the past. We present a processing pipeline to identify spatially and temporally variable fluctuations occurring in the first interferometer installed on a continuously rotating clinical CT gantry. The correlations of the vibrations in the modular grating setup are exploited to identify a small number of relevant vibration modes, allowing for an artifact-free reconstruction of a sample.

Based on well-established X-ray physics in computed tomography (CT) imaging, the spectral responses of different materials contained in lesions are different, which brings richer contrast information at various energy bins. Hence, obtaining the material decomposition of different tissue types and exploring its spectral information for lesion diagnosis becomes extremely valuable. The lungs are housed within the torso and consist of three natural materials, i.e., soft tissue, bone, and lung tissue. To benefit the lung nodule differentiation, this study innovatively proposed to use lung tissue as one basis material along with soft tissue and bone. This set of basis materials will yield a more accurate composition analysis of lung nodules and benefit the following differentiation. Moreover, a corresponding machine learning (ML)-based computer-aided diagnosis framework for lung nodule classification is also proposed and used for evaluation. Experimental results show the advantages of the virtual monoenergetic images (VMIs) generated with lung tissue material over the VMIs without lung tissue and conventional CT images in differentiating the malignancy from benign lung nodules. The gain of 9.63% in area under the receiver operating characteristic curve (AUC) scores indicated that the energy-enhanced tissue features from lung tissue have a great potential to improve lung nodule diagnosis performance.

Phase-contrast CT (PCCT) is an emerging tool that has found numerous applications, including applications to preclinical imaging. There remains a need for reducing the imaging time in current PCCT. One approach to reducing imaging time is to reduce the scanning angular range in PCCT. However, accurate image reconstruction from data collected over a limited angular range (LAR) is challenging because it poses a problem of accurate inversion of the PCCT imaging model that can be highly ill-conditioned in LAR scans. In this work, we conduct an investigation of accurate image reconstruction through inverting the imaging model for LAR scanning configurations in propagation-based (PB) PCCT. We have developed a directional-total-variation (DTV) algorithm for image reconstruction from knowledge of the discrete X-ray transform (DXT) over a LAR for CT imaging. Observing the mathematical similarity between the DXT in CT and the imaging model in PB-PCCT, we develop and tailor the DTV algorithm for image reconstruction from LAR data in PB-PCCT. Results of our study show that the tailored DTV algorithm can yield image reconstruction with reduced LAR artifacts that can be observed otherwise in images reconstructed by use of the existing algorithm in PB-PCCT imaging. For a given LAR, it can be divided into sub arcs of LARs. We also investigate a scanning configuration with two orthogonal arcs of LARs separated by 90° , and observe that the two-orthogonal-arc scanning configuration may allow image reconstruction more accurately than does a single-arc scanning configuration even though the total angular ranges in both scanning configurations are identical. While boundary images can be reconstructed from data, we develop the DTV algorithm for reconstruction of the image, i.e., the refractive index distribution, instead of its boundary image from data in PB-PCCT. Once the image is obtained, the Laplacian operator can be applied to it for yielding its boundary image.

Supervised deep convolutional neural network (CNN)-based methods have been actively used in clinical CT to reduce image noise. The networks of these methods are typically trained using paired high- and low-quality data from a large number of patients and/or phantom images. This training process is tedious, and the network trained under a given condition may not be generalizable to patient images acquired and reconstructed at different conditions. In this paper, we propose a self-trained deep CNN (ST_CNN) method which does not rely on pre-existing training datasets. The training is accomplished using extensive data augmentation through projection domain and the inference is applied to the data itself. Preliminary evaluation on patient images demonstrated that the proposed method could achieve similar image quality in comparison with conventional deep CNN denoising methods pre-trained on external datasets.

The study of rigid body dynamics in soft tissue is an essential part of orthopaedic imaging. Our focus is the foot and ankle structure, which consists of 28 bones surrounded by a variety of soft tissue such as 30 muscles, numerous ligaments, bursae and nerves. The importance of understanding the involved dynamics is evident by the frequent need to find functional joint replacements to fight diseases such as hindfoot arthritis. We study a simplified problem by constructing a phantom with two ”bones” submerged in silicone and a remotely controlled LEGO robot that moves one of the two bones. We perform motion registration by manipulating the bone segmented from an initial static scan and matching its digitally reconstructed radiographs to a sequence of scanned X-ray projections. The registration of the 3D volume to 2D projections requires the right combination of optimization and interpolation to achieve an optimal result. We test four different approaches with the help of the Flexible Algorithms for Image Registration (FAIR) toolbox.

Motion of gas bubbles (gastrointestinal gas) in abdominal region can produce significant artifacts during on-board CBCT scanning, which adversely affects the imaging quality and limits the process of CBCT-based adaptive planning and image-guided radiotherapy. In this study, we tested the effectiveness of simultaneous motion estimation and image reconstruction technique (SMRIE) for improving CBCT image quality and HU accuracy of abdominal scan. The improved image quality in the simulation study demonstrated that SMRIE technique is promising for on-board CBCT gas bubble motion artifact reduction.

In this work, we compare one-step and two-step approaches for X-ray computed tomography (CT) scatter correction and density reconstruction. X-ray CT is an important imaging technique in medical and industrial applications. In many cases, the presence of scattered X-rays leads to loss of contrast and undesirable artifacts in reconstructed images. Many approaches to computationally removing scatter treat scatter correction as a preprocessing step that is followed by a reconstruction step. Treating scatter correction and reconstruction jointly as a single, more complicated optimization problem is less studied. It is not clear from the existing literature how these two approaches compare in terms of reconstruction accuracy. In this paper, we compare idealized versions of these two approaches with synthetic experiments. Our results show that the one-step approach can offer improved reconstructions over the two-step approach, although the gap between them is highly object-dependent.

Dual-kV is a spectral CT modality that is currently available clinically, which traditionally employs energy integrating detectors, and in some cases acquire spectral data that is unregistered. A spectral method developed for photon-counting detectors, the constrained one-step spectral CT image reconstruction method (cOSSCIR), is designed to perform material decomposition directly from a set of spectral measurements. As such, the spectral sinograms need not be registered. Within the framework of cOSSCIR, dual-kV data is just two-window spectral CT. Furthermore, reconstructing unregistered data, such as that acquired by fast kV switching and slow kV switching systems, provides a potential extension of sOSSCIR to reconstructing clinical data. In this investigation we demonstrate cOSSCIR on unregistered dual-kV protocols using simulations and experimental phantom studies. Our results suggest that cOSSCIR can accurately recover the basis material maps using slow- and rapid-kV data compared to fully registered reconstructions.

Patient-based CT phantoms, with realistic image texture and densities, are essential tools for assessing and verifying CT performance in clinical practice. This study extends our previously presented 3D printing solution (PixelPrint) to patient-based phantoms with soft tissue and bone structures. To expand the Hounsfield Unit (HUs) range, we utilize a stone-based filament. Applying PixelPrint, we converted patient DICOM images directly into FDM printer instructions (G-code). Density was modeled as the ratio of filament to voxel volume to emulate attenuation profiles for each voxel, with the filament ratio controlled through continuous modification of the printing speed. Two different phantoms were designed to demonstrate the high reproducibility of our approach with micro-CT acquisitions, and to determine the mapping between filament line widths and HU values on a clinical CT system. Moreover, a third phantom based on a clinical cervical spine scan was manufactured and scanned with a clinical spectral CT scanner. CT image of the patient-based phantom closely resembles the original CT image both in texture and contrast levels. Measured differences between patient and phantom are around 10 HU for bone marrow voxels and around 150 HU for cortical bone. In addition, stone-based filament can accurately represent boney tissue structures across the different x-ray energies, as measured by spectral CT. This study demonstrates the feasibility of our 3D-printed patient-based phantoms to be extended to soft-tissue and bone structure while maintaining accurate organ geometry, image texture, and attenuation profiles for spectral CT.

Non-circular orbits in cone-beam CT (CBCT) imaging are increasingly being studied for potential benefits in field-of-view, dose reduction, improved image quality, minimal interference in guided procedures, metal artifact reduction, and more. While modern imaging systems such as robotic C-arms are enabling more freedom in potential orbit designs, practical implementation on such clinical systems remains challenging due to obstacles in critical stages of the workflow, including orbit realization, geometric calibration, and reconstruction. In this work, we build upon previous successes in clinical implementation and address key challenges in the geometric calibration stage with a novel calibration method. The resulting workflow eliminates the need for prior patient scans or dedicated calibration phantoms, and can be conducted in clinically relevant processing times.

X-ray photon-counting detector (PCD) offers low noise, high resolution, and spectral characterization, representing a next generation of CT and enabling new biomedical applications. It is well known that involuntary patient motion may induce image artifacts with conventional CT scanning, and this problem becomes more serious with PCD due to its high detector pitch and extended scan time. Furthermore, PCD often comes with a substantial number of bad pixels, making analytic image reconstruction challenging and ruling out state-of-the-art motion correction methods that are based on analytical reconstruction. In this paper, we extend our previous locally linear embedding (LLE) cone-beam motion correction method to the helical scanning geometry, which is especially desirable given the high cost of large-area PCD. In addition to our adaption of LLE-based parametric searching to helical cone-beam photon-counting CT geometry, we introduce an unreliable-volume mask to improve the motion estimation accuracy and perform incremental updating on gradually refined sampling grids for optimization of both accuracy and efficiency. Our numerical results demonstrate that our method reduces the estimation errors near the two longitudinal ends of the reconstructed volume and overall image quality. The experimental results on clinical photon-counting scans of the patient extremities show significant resolution improvement after motion correction using our method, which reveals subtle fine structures previously hidden under motion blurring and artifacts.

The x-ray computed tomography (CT) images with sparse-view data acquisition contain severe angular aliasing artifacts. The common denoising filters do not work well. The state-of-the-art methods to process the sparse-view CT images are deep learning based; they require a large amount of training data pairs. This paper considers a situation where no training data sets are available. All we have is one sparse scan of a patient. This paper attempts to use a BM3D filter to reduce the artifacts by introducing an artifact power spectral density function, which is calculated with computer simulations. The results in this paper show that the proposed method is not effective enough for practice applications. However, some insights may lead us to further investigations.

Fusion of X-ray projection images obtained with different exposure levels is a promising technique for studying objects with features beyond the dynamic range of the X-ray detector. Various multi-exposure fusion techniques are described in the literature, yet a direct comparison between these methods is not available. This was mainly due to the absence of objective quality measures dedicated to multi-exposure X-ray images and tomographic reconstructions, a problem remaining unsolved to this day. Therefore, in this work, we compare several fusion algorithms in terms of perceptual quality using recently reported quality measures based on structural similarity. Moreover, we investigate whether these quality measures apply to tomographic slices as well. Our results indicate that the reliability of the quality measures is more convincing for fused projection images as opposed to reconstructed slices. Additionally, it is shown that fusion algorithms developed for optical photography are also suitable for multi-exposure X-ray image fusion to increase perceptual quality.

We present a geometric calibration method for integrating a seven degrees of freedom robotic arm as a sample holder within an existing laboratory X-ray computed tomography setup. We aim to provide a flexible sample holder that is able to execute non-standard and task-specific trajectories for complex samples. The calibration is necessary to identify the accurate pose of the sample which deviates from the expected pose due to inaccurate placement of the robotic arm. The robotic arm is integrated with a unified software package that allows for path planning, collision detection, geometric calibration and reconstruction of the sample. With our software the user is able to command the robotic arm to execute arbitrary trajectories for a given sample in a safe manner and output its reconstruction to the user. We present experimental results with a circular trajectory where the robotic sample holder achieves identical visual quality compared to a conventional sample holder.

Photon counting detectors (PCDs) with energy discrimination capabilities allow us to perform quantitative material decomposition with high spatial resolution. Although PCDs provide more spectral information than conventional energy integrating detectors (EIDs), it is more challenging for the system to transmit projection data from the detectors across the slip ring to the processing computer and store the data, due to the increased amount of data with increasing number of energy bins. To address this problem, many approaches have been proposed to compress the bin data while maintaining the image quality. In this work, we compare the performance of strategies to reduce projection data and determine the optimal choice of bin compression strategies and the number of measurements for multiple tasks. We first obtain the optimal thresholds for conventional energy bins, as determined by minimizing the Cramér–Rao lower bound (CRLB) for material decomposition tasks with a realistic silicon detector energy response. We then consider the case of reducing data from eight native energy bins by forming weighted sums, either with binary weights or continuous weights, by minimizing the relative CRLB between the compressed measurements and the original eight bins. We then evaluate their respective performance using Monte Carlo simulation for a head phantom. The results show that the continuous weights strategy is superior to others, with low bias and less than 10% variance penalty for two weighted sums, with a data reduction of 75% within a large material thickness space. The other strategies have up to 50% variance penalty compared with the original eight bins and are less robust when there is photon starvation. With additional weighted measurements, the continuous weights method can achieve less than 1% variance penalty when reducing the eight native energy bins to half the number of measurements. Overall, combining energy bins by forming weighted sums with continuous weights is an effective strategy for reducing data while preserving spectral information.

The directional total variation algorithm (DTV) reported in the literature achieves promising results for limited angle reconstructions, especially when the scanning angular range is very small. However, the visible edge prior for limited-angle CT is not explicitly considered by DTV. In this paper, a variant of the DTV model is proposed which explicitly builds into the visible edge prior developed by Quinto et al. Numerical experiments show that the proposed model and algorithm produce very competitive results compared to DTV.

Deep learning (DL) are being extensively investigated for low-dose computed tomography (CT). The success of DL lies in the availability of big data, learning the non-linear mapping of low-dose CT to target images based on convolutional neural networks. However, due to the commercial confidentiality of CT vendors, there are very few publicly raw projection data available to simulate paired training data, which greatly reduces the generalization and performance of the network. In the paper, we propose a dual-task learning network (DTNet) for low-dose CT simulation and denoising at arbitrary dose levels simultaneously. The DTNet can integrate low-lose CT simulation and denoising into a unified optimization framework by learning the joint distribution of low-dose CT and normal-dose CT data. Specifically, in the simulation task, we propose to train the simulation network by learning a mapping from normal-dose to low-dose at different levels, where the dose level can be continuously controlled by a noise factor. In the denoising task, we propose a multi-level low-dose CT learning strategy to train the denoising network, learning many-to-one mapping. The experimental results demonstrate the effectiveness of our proposed method in low-dose CT simulation and denoising at arbitrary dose levels.

Photon-counting computed tomography (PCCT) can simultaneously obtain multi-energy data with abundant energy-dependent material-specific information of the scanned object. However, the photon counts in each energy bin are decreased and the collected data suffers from photon starvation effects, which degrades the quality of the reconstructed PCCT images. To solve it, many statistical iteration reconstruction (SIR) methods have been proposed by constructing data-fidelity and prior information terms to suppress noise and remove artifacts. However, most of the current SIR methods assume the noise in PCCT images follows a Gaussian distribution, which deviates the real distribution of the noise in PCCT images. Therefore, we propose a new statistical iteration reconstruction method by considering more complex noise distribution in reality. Specifically, Gaussian Mixture Model (GMM), which is a universal approximator for any continuous density function, is utilized to model the noise in PCCT images. Moreover, the multi-energy PCCT images are treated as a 3- order tensor which is regularized by three dimensional total variation (3DTV) prior term. Finally, a statistical iteration reconstruction model based on GMM and 3DTV is established for PCCT imaging. For shorten, we call the presented reconstruction model as “GMM-3DTV”. We then develop an expectation-maximization (EM) algorithm to solve the presented GMM-3DTV method. Numerical studies demonstrate the improvements of the presented GMM-3DTV method over the competing methods.

Photon-counting spectral CT is a novel technology with a lot of promise. However, one common issue is detector inhomogeneity which results in streak artifacts in the sinogram domain and ring artifacts in the image domain. These rings are very conspicuous and limit the clinical usefulness of the images. We propose a deep learning based image processing technique for ring artifact correction in the sinogram domain. In particular, we train a UNet using a perceptual loss function with VGG16 as feature extractor to remove streak artifacts in the basis sinograms. Our results show that this method can successfully produce ring-corrected virtual monoenergetic images at a range of energy levels.

We reported our developments of a photon counting detector (PCD) based multi-energy cone beam CT (ME-CBCT) for preclinical small animal radiation research. The development was based on an existing preclinical small animal irradiator that includes a gantry mounted x-ray tube for imaging and radiation experiments in a self-shielded cabinet. We installed a PCD on the gantry and developed the data acquisition, processing, and image reconstruction pipeline to reconstruct CBCT images at three energy channels. We determined the optimal energy thresholds as 26, 56, and 90 keV to achieve uniform signal-to-noise ratio among energy channels. Pixel-based detector response calibration was performed to remove ring artifacts in the reconstructed CBCT images. The average difference between measured x-ray attenuation coefficients of targeted materials from ME-CBCT images and analytically calculated values was 10%. We decomposed the ME-CBCT images into images of water and bone material via an optimization model. The PCD-based ME-CBCT is expected to facilitate critical tasks in preclinical small animal irradiation researches, such as improved accuracy of radiation dose calculations in experiment planning.

Deep learning (DL) shows promise of advantages over conventional signal processing techniques in a variety of imaging applications. The networks’ being trained from examples of data rather than explicitly designed allows them to learn signal and noise characteristics to most effectively construct a mapping from corrupted data to higher quality representations. In inverse problems, one has options of applying DL in the domain of the originally captured data, in the transformed domain of the desired final representation, or both. X-ray computed tomography (CT), one of the most valuable tools in medical diagnostics, is already being improved by DL methods. Whether for removal of common quantum noise resulting from the Poisson-distributed photon counts, or for reduction of the ill effects of metal implants on image quality, researchers have begun employing DL widely in CT. The selection of training data is driven quite directly by the corruption on which the focus lies. However, the way in which differences between the target signal and measured data is penalized in training generally follows conventional, pointwise loss functions. This work introduces a creative technique for favoring reconstruction characteristics that are not well described by norms such as mean-squared or mean-absolute error. Particularly in a field such as X-ray CT, where radiologists’ subjective preferences in image characteristics are key to acceptance, it may be desirable to penalize differences in DL more creatively. This penalty may be applied in the data domain, here the CT sinogram, or in the reconstructed image. We design loss functions for both shaping and selectively preserving frequency content of the signal.

Accurate models of the X-ray attenuation process are required for quantitative estimation of iodine concentration with model-based reconstruction methods. The choice of model is influenced not only by the accuracy sought but also by the increasing complexity when more free parameters need to be reconstructed. The applicability of three attenuation models was investigated in a single pixel problem using either two or three monochromatic beams near the K-edge energy of iodine. We found that an empirical model with five components, proposed by Midgley, leads to the lowest error when modeling iodine free materials and small error in estimating iodine concentration (0.1% and 3.39%), whereas the decomposition into contributions due to photoelectric effect and incoherent scatter results in more accurate estimation of the iodine concentration (0.72%) but has larger error (8.9%) when reconstructing iodine free materials. Decomposition into base materials shows the worst results on both objectives (8.9% and 62%).

Motion during CT scans causes artifacts that can severely degrade the image quality. We propose a motion compensation algorithm that can be combined with reconstruction algorithms that contain a rebinning step like the weighted filtered backprojection algorithm. Therefore, we assume that the motion present during acquisition is known, and we extend the backprojection step in the reconstruction to consider this motion, reducing the motion artifacts in the resulting images. Furthermore, we propose a combination of our motion compensation algorithm with two versions of an iterative weighted filtered backprojection algorithm.

In this work related to the use of a commercial medical CT scanner for the non-destructive analysis of highly attenuating materials (mineral samples), the effect of backprojection techniques and truncation artifacts corrections were explored. For small ROIs, the CT couch interferes significantly in images of small samples (few centimeters). An iterative reconstruction algorithm (OSC-TV) was used to perform reconstructions from uncorrected raw projection data made available through a collaboration with the CT vendor, who provided binaries and methods to remove low-level, proprietary data corrections (for beam hardening). The OSC-TV algorithm is customizable, allowing for the use of different forward-projection and backprojection techniques. Reconstruction parameters were tuned by performing simulations in a virtual phantom involving highly attenuating materials. Strategies to reconstruct small ROIs were also explored, with the objective of reducing truncation artifacts. Three samples were scanned to compare a ray-driven backprojection and a voxel-driven backprojection technique based on bilinear interpolation. The voxel-driven approach led to better results in terms of noise and reconstruction artifacts. An iterative ROI reconstruction technique was used to reconstruct small ROIs. This technique allows obtaining a sinogram with the projections of the ROI only. With that, truncation artifacts were reduced, which led to images with less blurring.

In interventional radiology, 3D reconstruction of devices such as needles would increase the precision of procedures. Doing so with CBCT is time-consuming and increases the X-ray dose. Needles being sparse, a compressed-sensing reconstruction approach seems viable. We thus investigate the interest of directional total variation as an adequate prior for anisotropic devices. We introduce a decomposition method that allows several a priori directions to be considered at once as well as excludes the anatomical background that is not sparse. The capacity of the method is illustrated on simulations of limited-angle acquisitions. It is shown to allow good reconstruction of the needles from a small angular coverage, even if the anatomical background cannot be recovered.

Tomosynthesis offers an alternative to planar radiography providing pseudo-tomographic information at a much lower radiation dose than CT. The fact that it cannot convey information about the density poses a major limitation towards the use of tomosynthesis in chest imaging, due to the wide range of pathologies that present an increase in the density of the pulmonary parenchyma. Previous works have attempted to improve image quality through enhanced analytical, iterative algorithms, or including a deep learning-based step in the reconstruction, but the results shown are still far from the quantitative information of a CT. In this work, we propose a reconstruction methodology consisting of a filtered back-projection step followed by post-processing based on Deep Learning to obtain a tomographic image closer to CT. Preliminary results show the potential of the proposed methodology to obtain true tomographic information from tomosynthesis data, which could replace CT scans in applications where the radiation dose is critical.

The efficacy of interventional treatments highly relies on an accurate identification of the target lesions and the interventional tools in the guidance images. Whereas X-ray radiography poses low doses to the patient, its weakness is in the superposition of the different image structures in a 2D image. Cone-beam computed tomography (CBCT) might look ideal providing exact 3D information, however this is at the cost of a higher radiation dose, longer imaging time, and more space requirements in the operating room. Introducing some depth information with relatively low dose, and requiring less space, digital tomosynthesis (DTS) is a potential candidate for guiding interventions. However, due to the few number of projections and to the limited angle acquisition, DTS has poor depth resolution. Since high quality patient-specific prior CT scans are usually performed prior to the intervention for diagnosis or to plan the intervention, and given that such images share a fair amount of information with the intraoperative DTS images, we propose in this work a prior-based iterative reconstruction framework to improve the intraoperative DTS image quality. The framework is based on registering the prior CT image to an intermediate low-quality intraoperative DTS image, then iteratively re-reconstructing the intraoperative DTS image using the co-registered prior CT as the starting image. We acquired prior CT and intraoperative CBCT data of a liver phantom and simulated some intraoperative DTS projection images using a spherical ellipse scan geometry. Our results show a great improvement in the DTS image quality with the proposed method and prove the importance of choosing a good starting point for the iterative DTS reconstruction.

Since X-ray scattering is a major cause of artifacts, its correction is a crucial step in almost any CT application. Most existing approaches, however, are based on complex theoretical models that need to be tailored to that particular application. To perform scatter estimation in absence of such models, we propose the unsupervised deep scatter estimation (uDSE). Here, uDSE combines a scatter estimation network that operates in projection domain with a scatter correction layer and CT reconstruction layer. In that way scatter estimation can be trained using an unsupervised Wassersten GAN (WGAN) setup in which the parameters of the scatter estimation network are optimized such that the resulting scatter corrected reconstructions cannot be distinguished from samples of a true artifact-free reference set. To demonstrate the feasibility of the proposed approach, uDSE is evaluated for simulated CBCT scans. Applied to the corresponding test data, uDSE is able to remove most of the present scatter artifacts and yields similar CT value accuracy (mean error of 27.9 HU vs. 24.7 HU) as a state-of-the-art supervised scatter estimation approach. Thus, uDSE may be used in the future to learn scatter estimation in cases where labels are not available or cannot be generated with sufficient accuracy.

This paper presents some practical realizations of image reconstruction methods for spiral cone-beam tomography scanners in which an X-ray tube with a flying focal spot is used. These methods are related to the original formulated 3D statistical model-based iterative reconstruction approach for tomography with flying focal spot. The conception proposed here is based on principles of a statistical model-based iterative reconstruction (MBIR) methodology, where the reconstruction problem is formulated as a shift-invariant system (a continuous-to-continuous data model). We adopted nutating reconstruction-based approaches, i.e. the advanced single slice rebinning methodology (usually applied in CT scanners with X-ray tubes with a flying focal spot), and a procedure compliant with the FDK scheme. We showed that our methods significantly improve the quality of obtained images compared to the traditional FBP algorithms. Consequently, it can allow for a reduction in the x-ray dose absorbed by a patient. Additionally, we show that our approach can be competitive in terms of the time of calculations, especially if we consider commercially used statistical reconstruction systems.

Imaging systems require to be calibrated. The geometric calibration consists in estimating several parameters describing the projection geometry. Just like in computer vision for cameras, intrinsic parameters characterize the internal parameters of X-ray projection system. The extrinsic parameters define the orientation and position of the acquisition system. In X-ray computed tomography (CT), the acquisition systems are generally composed of a detector and a X-ray source. The object to be reconstructed lies in-between. A perfect knowledge of the calibration parameters is needed for the reconstruction algorithm to reduce artefacts. In this paper, we focus on off-line calibration methods for 1D linear X-ray detector systems. We first introduce a calibration method for systems composed of a single linear detector. This method solves the problem of calibration in two steps using calibration objects based on four co-planar lines. Moreover, we generalize the single linear detector geometric calibration method to a multi-linear detector system. We compare four different numerical models and methods. Three are based on non-linear equation systems. Finally, we propose an adaptative calibration object.

Liver vessel segmentation is important in diagnosing and treating liver diseases. Iodine-based contrast agents are typically used to improve liver vessel segmentation by enhancing vascular structure contrast. However, conventional computed tomography (CT) is still limited with low contrast due to energy-integrating detectors. Photon counting detector-based computed tomography (PCD-CT) shows the high vascular structure contrast in CT images using multi-energy information, thereby allowing accurate liver vessel segmentation. In this paper, we propose a deep learning-based liver vessel segmentation method which takes advantages of the multi-energy information from PCD-CT. We develop a 3D UNet to segment vascular structures within the liver from 4 multi-energy bin images which separates iodine contrast agents. The experimental results on simulated abdominal phantom dataset demonstrated that our proposed method for the PCD-CT outperformed the standard deep learning segmentation method with conventional CT in terms of dice overlap score and 3D vascular structure visualization.

Metal artifacts are one of the most common reasons for reduced image quality and usability in polychromatic cone-beam CT. In this work, we revisit empirical beam hardening correction algorithm and propose a few practical optimizations to simplify its application. First, fuzzy C-means segmentation method is used to perform an automatic segmentation of the metal component. Second, a minimum variance optimization technique provides a suitable combination of correction basis images. Finally, a sub-volume (spatially varying) optimization method is used to account for a varying contribution of metal artifacts through the image. We apply the modified algorithm to datasets from cone-beam CT and evaluate its performance.

Deep learning-based image denoising and reconstruction methods have shown promising results for low-dose CT. When high-quality reference images are not available for training the network, researchers found a powerful and effective counterpart called Noise2Noise, which trains the neural network using paired data with independent noise. However, it is uncommon to have paired CT scans with independent noise (e.g., from two scans). In this paper, a method is proposed to generate such paired data for potential usage in deep learning training by simultaneously simulating a low-dose image at arbitrary dose level and an image with independent noise from a single CT scan. Their independence is investigated both analytically and numerically. In our numerical study, a Shepp-Logan phantom was utilized in MATLAB to generate the ground-truth, normal-dose, and low-dose images for reference. Noise images were obtained for analysis by subtracting the ground-truth from the noisy images, including the normal-dose/low-dose images and the paired products of our proposed method. Our numerical study matches the analytical results very well, showing that the paired images are not correlated. Under an additional assumption that they form a bivariate normal distribution, they are also independent. The proposed method can produce a series of paired images at arbitrary dose level given one CT scan, which provides a powerful new method to enrich the diversity of low-dose data for deep learning.

Augmentation strategies have been suggested to overcome issues related to limited training data for the Deep Learning (DL)-based CT image denoising problem. Although augmentation is, indeed, a good machine learning practice, the extent of improvements achieved by DL-based CT denoising solvers through augmentation pipelines remains to be quantified. Accordingly, in this work, we make use of two different deep neural networks (the REDCNN, the DnCNN) to quantify gains in CT image quality through augmentation. The augmentation strategies considered include computer vision inspired strategies (like scaling, rotating, flipping, image-blending) and the CT forward model-based noise (radiation dose) insertion. Likewise, image qualities considered in this work include visual perception- and data fidelity-based global metrics (like the PSNR, the SSIM, the RSE) which are common in the computer vision literature, and CT bench tests (like the NPS, the MTF) and a task-based detectability assessment (the LCD) from the CT imaging literature. Our preliminary results indicate that the DL solvers trained on augmented data show gains in the global metrics, in low-frequency noise texture components and in the MTF values as compared to the ones trained on non-augmented data. However, when the augmented DL-solvers were compared against their low-dose counterparts, their performance - with respect to the noise frequency components, resolution, and detection task - was not all improved, and in some cases even worse

Photon-counting detector based spectral computed tomography (CT) has great potential in material decomposition, tissue characterization, lesion detection, and other applications. For a fixed total photon number or radiation dose, the increase of channel number will lead to the decrease of the photon number in each channel, resulting in degraded image quality of the reconstructed image. This is difficult to meet the practical applications for material decomposition, tissue characterization, lesion detection, etc. To improve the quality of image reconstruction, we propose a spectral CT reconstruction algorithm based on joint multi-channel total generalized variation (TGV) minimization and tensor decomposition. On one hand, the algorithm takes joint multi-channel TGV function as regularization. The total generalized variation is extended to the vector, and the sparsity of singular value is used to promote the linear dependence of the image gradient. On the other hand, the multi-channel images share the same physical structure, and the algorithm employs the non-local feature similarity in the image domain. Similar image blocks are clustered into a four-order tensor group, and the noise was reduced by sparse representation of high-dimensional tensors. Experiment results show the proposed algorithm can further improve the quality of reconstructed image and preserve the edge and details of the spectral CT image.