This PDF file contains the front matter associated with SPIE Proceedings Volume 12286, including the Title Page, Copyright information, Table of Contents, and Conference Committee listings.

A realistic 3D anthropomorphic software model of microcalcifications may serve as a useful tool to assess the performance of breast imaging applications through simulations. We present a method allowing to simulate visually realistic microcalcifications with large morphological variability. Principal component analysis (PCA) was used to analyze the shape of 281 biopsied microcalcifications imaged with a micro-CT. The PCA analysis requires the same number of shape components for each input microcalcification. Therefore, the voxel-based microcalcifications were converted to a surface mesh with same number of vertices using a marching cube algorithm. The vertices were registered using an iterative closest point algorithm and a simulated annealing algorithm. To evaluate the approach, input microcalcifications were reconstructed by progressively adding principal components. Input and reconstructed microcalcifications were visually and quantitatively compared. New microcalcifications were simulated using randomly sampled principal components determined from the PCA applied to the input microcalcifications, and their realism was appreciated through visual assessment. Preliminary results have shown that input microcalcifications can be reconstructed with high visual fidelity when using 62 principal components, representing 99.5% variance. For that condition, the average L2 norm and dice coefficient were respectively 10.5 μm and 0.93. Newly generated microcalcifications with 62 principal components were found to be visually similar, while not identical, to input microcalcifications. The proposed PCA model of microcalcification shapes allows to successfully reconstruct input microcalcifications and to generate new visually realistic microcalcifications with various morphologies.

Many works have investigated methods to assess the quality of mammography images using objective image quality metrics. However, few studies have evaluated the ability of these metrics to predict the performance of human observers on specific tasks related to mammographic examination that are highly dependent on image quality. The propose of this work is to evaluate the quality of digital mammography acquired at a range of radiation doses through a set of objective metrics and to compare the results with the performance of human observers in the task of locating microcalcification clusters in these images. A dataset of 100 synthetic mammograms was simulated using a virtual clinical trials software. Microcalcification clusters of different sizes and contrasts were computationally inserted into the images. Acquisitions with five different radiation doses were simulated using a noise injection method proposed in a previous work. Four medical physicists with experience in analysis of mammographic images participated in the microcalcification cluster localization tests. The quality of digital mammography images was assessed considering nine well-known objective metrics. The metrics were calculated on both the raw data (DICOM ‘for processing’ tag) and the processed images (DICOM ‘for presentation’ tag). Finally, the association between readers performance and image quality index was conducted by calculating the percentage variation of all metrics as a function of radiation dose, taking the standard dose as a reference. Although the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) are the most used in the literature, our results showed that Quality Index based on Local Variance (QILV) is the objective metric that best describes the behavior of human visual perception with the variation of radiation dose in digital mammography.

PURPOSE: To investigate differences in microcalcification detection performance for different acquisition setups in digital breast tomosynthesis (DBT), a convex dose distribution and sparser number of projections compared to the standard set-up was evaluated via a virtual clinical trial (VCT). METHODS AND MATERIALS: Following the Institutional Review Board (IRB) approval and patient consent, mediolateral oblique (MLO) DBT views were acquired at twice the automatic exposure controlled (AEC) dose level; omitting the craniocaudal (CC) view limited the total examination dose. Microcalcification clusters were simulated into the DBT projections and noise was added to simulate lower dose levels. Three set-ups were evaluated: (1) 25 DBT projections acquired with a fixed dose/projection at the clinically used AEC dose level, (2) 25 DBT projections with dose/projection following a convex dose distribution along the scan arc, and (3) 13 DBT projections at higher dose with the total scan dose equal to the AEC dose level and preserving the angular range of 50° (sparse). For the convex set-up, dose/projection started at 0.035 mGy at the extremes and increased to 0.163 mGy for the central projection. A Siemens prototype algorithm was used for reconstruction. An alternative free-response receiver operating characteristic (AFROC) study was conducted with 6 readers to compare the microcalcification detection between the acquisition set-ups. Sixty cropped VOIs of 50x50x(breast thickness) mm3 per set-up were included, of which 50% contained a microcalcification cluster. In addition to localization of the cluster, the readers were asked to count the individual calcifications. The area under the AFROC curve was used to compare the different acquisition set-ups and a paired t-test was used to test significance. RESULTS: The AUCs for the standard, convex and sparse set-up were 0.97±0.01, 0.95±0.02 and 0.89±0.03, respectively, indicating no significant difference between standard and convex set-up (p=0.309), but a significant decrease in detectability was found for the sparse set-up (p=0.001). The number of detected calcifications per cluster was not significantly different between standard and convex set-ups (p=0.049), with 42%±9% and 40%±8%, respectively. The sparse set-up scored lower with a relative number of detected microcalcifications of 34%±11%, but this decrease was not significant (p=0.031). CONCLUSION: A convex dose distribution that increased dose along the scan arc towards the central projections did not increase detectability of microcalcifications in the DBT planes compared to the current AEC set-up. Conversely, a sparse set of projections acquired over the total scan arc decreased microcalcification detectability compared to the variable dose and current clinical set-up.

This work proposes an empirical model for tuning spatial resolution and noise in simulated images in virtual clinical trials in x-ray breast imaging. In extending previous studies performed for direct conversion a-Se detectors used in digital mammography and digital breast tomosynthesis, this work introduces the model for the case of cone-beam computed tomography dedicated to the breast that uses a indirect conversion flat-panel detector. In the simulations, the detector is modeled as an absorbing layer whose material and thickness reflect those of the scintillator of the detector of a clinical scanner. The simulated images are then computed as a dose deposit map. The detector response curve, modulation transfer function (MTF) and noise power spectrum (NPS) were measured on a real detector. The same measurements were replicated in-silico for the simulated detector and scanner. The comparison of simulated and measured detector response curves permits to recover pixel values at the clinical scale. The difference between the simulated and measured MTFs permitted to introduce a linear filter for compensating simulated model simplification that determines a better spatial resolution in the simulated images with respect to real images. This filter presented a Gaussian shape in the Fourier domain with a standard deviation of 1.09 mm^-1 , derived from those of the measured and simulated MTF curves, of 0.86 mm^-1 and 1.41 mm^-1 , respectively. Finally, the analysis of the NPS permits to compensate for noise characteristics due to the simulated model simplifications. The model applied to the simulated projection images produced MTF and normalized NPS in simulated 3D images, comparable to those obtained for the clinical scanner.

Image quality directly influences the accuracy of lesion detection and characterization in x-ray mammograms. Thus, it is crucial that acceptable image quality is maintained while using as little ionizing radiation as possible. In this scenario, denoising plays an important role in recovering image quality while keeping constant radiation dose. Although most ‘off-the-shelf’ denoising algorithms assume signal-independent and frequency-independent (white) Gaussian noise, in x-ray generation and detection this assumption is seldom valid. In this work we leverage a recently published variance-stabilizing transform and a frequency-dependent denoising algorithm to address signal-dependent and frequency-dependent denoising of x-ray mammograms subject to structural and correlated noise. To illustrate the application of the proposed pipeline, we restored synthetic mammograms generated by a virtual clinical trial platform. The results showed that the denoising pipeline was able to recover the quality of mammograms acquired at lower radiation levels to achieve similar image quality of full-dose acquisitions, in terms of the QILV, residual variance and power spectrum metrics. The bias² metric indicates that even though the pipeline is able to achieve very similar noise levels to a full-dose acquisition, there is a penalty to the signal, which becomes biased due to blur and smearing as the dose level is reduced.

We have developed a method for simultaneous tomosynthesis and mechanical imaging, called DBTMI. Mechanical imaging measures the stress distribution over the compressed breast surface. Malignant tissue is usually stiffer than benign, which results in higher stress on the compressed breast and enables to distinguish malignant from benign findings. By combining tomosynthesis and mechanical imaging, we could improve cancer detection accuracy by reducing the number of false positive findings. In this study we have analysed clinical DBTMI data, collected from 52 women from an ongoing pilot study at the Skåne University Hospital, Malmö, Sweden. We measured the range of the average stress over the breast surface, the range of average stress over the location of suspected lesions, and the normalized stress over the lesion location. Preliminary results show that the range of stress over the breast surface was 1.23-5.84 kPa, the range over the lesion location 2.10-10.10 kPa, and the normalized stress 1.12-2.44 over the lesion location. Overall, the local stress over malignant lesions was higher than the average stress over the entire breast surface. This is the first step investigating criteria to distinguish between malignant and benign findings based upon clinical DBTMI data.

Dedicated breast-imaging scanners using radiotracers, e.g., PET scanners (positron emission tomography) have been proposed and evaluated since the late 1980s [1]. These systems trade a reduction in the size of the imaging field of view for improved resolution, and potentially also a lower cost, higher sensitivity, and a smaller form factor. The higher resolution can improve both detection and quantitation of concentrations of radiotracer, although for the latter to be true, tomographic imaging is mandatory. Several commercial dedicated breast PEM (positron emission mammography), PET scanners, and gamma camera systems have been developed and marketed [2].

Contrast-enhanced digital mammography (CEDM) is used to detect iodine uptake in breast lesions. Iodine concentrations inside or around breast lesions could be used as a biomarker, provided a properly characterized quantification method is implemented. In this work, we have evaluated a method to quantify iodine concentrations in CEDM in terms of its intrinsic linearity, bias and variability. This evaluation was performed in a virtual clinical trial (VCT) environment, simulating anthropomorphic breast phantoms containing solid and liquid lesions with different iodine concentrations. Our results showed that anatomical variables such as breast size and lesion size and composition have a considerable effect on the iodine quantification. The method was linear in the clinical iodine concentration range, and showed an approximately constant 1 mg/cm² bias in the 0 – 2 mg/cm² range for both solid and liquid lesions. Corrections were proposed that reduced the variability due to breast size, lesion size, and composition.

Background and purpose: The recent emergence of neural networks models for the analysis of breast images has been a breakthrough in computer aided diagnostic. This approach was not yet developed in Contrast Enhanced Spectral Mammography (CESM) where access to large databases is complex. This work proposes a deep-learning-based Computer Aided Diagnostic development for CESM recombined images able to detect lesions and classify cases. Material and methods: A large CESM diagnostic dataset with biopsy-proven lesions was collected from various hospitals and different acquisition systems. The annotated data were split on a patient level for the training (55%), validation (15%) and test (30%) of a deep neural network with a state-of-the-art detection architecture. Free Receiver Operating Characteristic (FROC) was used to evaluate the model for the detection of 1) all lesions, 2) biopsied lesions and 3) malignant lesions. ROC curve was used to evaluate breast cancer classification. The metrics were finally compared to clinical results. Results: For the evaluation of the malignant lesion detection, at high sensitivity (Se<0.95), the false positive rate was at 0.61 per image. For the classification of malignant cases, the model reached an Area Under the Curve (AUC) in the range of clinical CESM diagnostic results. Conclusion: This CAD is the first development of a lesion detection and classification model for CESM images. Trained on a large dataset, it has the potential to be used for helping the management of biopsy decision and for helping the radiologist detecting complex lesions that could modify the clinical treatment.

The purpose of this study is to investigate the prediction of Ki-67 expression of breast cancers using MRI parameters from ultrafast (UF) DCE-MRI, DWI, T2WI, and the lesion size. Breast MRI was performed with a 3T scanner using dedicated breast coils. UF DCE-MRI was obtained using Compressed Sensing-VIBE (prototype sequence). As a kinetic parameter of UF DCE-MRI, maximum slope (MS) was defined as percentage relative enhancement (%/s), and time to enhance (TTE) was defined as the time interval between the aorta and lesion enhancement. The apparent diffusion coefficient (ADC) was derived from DWI. Two radiologists measured each MR parameter, and inter-rater agreement was evaluated. Univariate and multivariate logistic regression analyses were perfomed to predict low Ki-67 (<; 14%) and high Ki-67 (≥ 14%) expression using MS, TTE, ADC, T2- signal intensity (SI), and lesion size. The significant parameters (p-values of < 0.05) were selected for the prediction model, and the diagnostic performance of the model was evaluated using ROC curve analysis. A total of 191 invasive carcinomas defined as mass lesions were included (72 low Ki-67/ 119 high Ki-67 lesions). The inter-rater agreements of all parameters were excellent. After univariate and multivariate logistic regression analysis, ADC and lesion size remained significant parameters. Using these significant parameters, the multi-parametric prediction model yielded an AUC of 0.77 (95%CI of 0.70-0.84) (sensitivity 72.3%, specificity 76.4%, and PPV 83.5%, and NPV 62.5%). DWI parameter (ADC) may be more valuable than UF DCE-MRI parameters (MS, TTE) to predict high Ki-67 in mass-shaped invasive breast carcinoma.

In multimodal diagnosis for early breast cancer detection, spatial alignment by means of image registration is an important task. We develop patient-specific biomechanical models of the breast, for which one of the challenges is automatic segmentation for magnetic resonance imaging (MRI) of the breast. In this paper, we propose a novel method using unsupervised neural networks with pre-processing and post-processing to enable automatic breast MRI segmentation for three tissue types simultaneously: fatty, glandular, and muscular tissue. Pre-processing aims at facilitating training of the network. The architecture of neural network is a Kanezaki-net extended to 3D and consists of two sub-networks. Post-processing is enhancing the obtained segmentations by removing common errors. 25 datasets of T2 weighted MRI from the Medical University of Vienna have been evaluated qualitatively by two observers while eight datasets have been evaluated quantitatively based on a ground truth annotated by a medical practitioner. As a result of the qualitative evaluation, 22 out of 25 are usable for biomechanical models. Quantitatively, we achieved an average dice coefficient of 0.88 for fatty tissue, 0.5 for glandular tissue, and 0.86 for muscular tissue. The proposed method can serve as a robust method for automatic generation of biomechanical models.

The importance of the three-dimensional (3D) pathological observation of biological soft tissues has increased in recent year, and various visualization tools to obtain 3D information easily and analysis methods focusing on the 3D micro structures have been developed. Refraction-contrast computed tomography based on x-ray dark-field imaging technique (XDFI) is one of the powerful methods with a high contrast and spatial resolution. In this study, in order to apply XDFI as new pathology tool, we will develop the x-ray optics and the x-ray camera, which are important components of the XDFI imaging system, to achieve a spatial resolution of 5 μm and evaluate the spatial resolution by experiments of the x-ray micro chart and the breast tissue specimen.

Feature-based registration algorithms can be used to establish spatial correspondence between two image. Therefore, anatomical landmarks such as the breast boundary, pectoral muscle, nipple, duct and vessels need to be considered. The aim of this paper is to introduce a new approach which combine the pectoral muscle segmentation and nipple location, considering mammography quality assumptions. Pectoral muscle is initialized as a straight line from the top of the image to the nipple level. Afterwards, both pectoral muscle boundary and nipple position are optimized using an iterative approach. The results show that the nipple is localized on the contour of the corresponding area (error smaller than 10 mm) while the Dice’s coefficient of the pectoral muscle segmentation is equal to 0.84 ± 0.12 using a straight line which is improved using a Chan-Vese active contour approach, reaching 0.87 ± 0.13. Our algorithm is easily generalized and portable to a different mammographic system since it barely depends on images statistics -i.e. pixel intensity values-, and is just based on geometrical considerations.

Early detection of breast cancer through mammographic screening can only be achieved with high quality mammograms. In this study an experienced radiologist and radiographer scored 127 mammographic screening exams with MLO and CC views of left and right breasts using 18 different positioning quality criteria. This subjective evaluation of the positioning quality was compared to the objective and automatic assessment by Volpara TruPGMI (Volpara Health, New Zealand). The quality criteria on missed tissue at medial or lateral side of the breast were in agreement with the software for the radiographer but was scored differently by the radiologist. The criterion on the nipple in profile showed good agreement between the readers and the software. The important criterion on the number of images that had to be repeated showed that even though the same amount of cases was rated to be repeated, the majority of the cases were discordant between radiologist and software, the agreement with the radiographer was better. The presence of folds in the pectoral muscle, the adequate depiction of the pectoral muscle and inframammary angle on MLO view showed an acceptable agreement between the readers and software. Finally, the overall positioning quality was rated as Perfect, Good, Moderate or Inadequate. The extreme ratings of Perfect and Inadequate showed high agreement between readers and software. However the number of intermediate ratings “Moderate” and “Good” were very different. For the readers the majority of the images was “Good” whereas the software scored most often “Moderate”. Subjective positioning quality monitoring is prone to high reader variability; this can be overcome via the use of automatic measurements with software. Nevertheless, prior to the use of automatic quality monitoring software in clinical practice, a careful evaluation and validation is needed.

Modelling of the breast surface shape under compression in the cranio-caudal and medio-lateral oblique views could advance the development of image processing techniques and of dosimetric estimates in digital mammography and digital breast tomosynthesis. Our goal is to compare the performance of a previously tested and used optical structured light scanning system (SLSS) capable of capturing the breast shape under compression to that of an infrared smartphone-based SLSS. Their performance was compared by scanning a cuboid phantom and two breast shaped phantoms (30 mm and 74 mm thick). Ten scans of the cuboid phantom were acquired with each scanner, and the measured length and thickness of the scanned shape were compared against the ground truth and between the two scanners. The performance of the scanners regarding breast-like phantoms was evaluated by calculating the maximum and mean distance, along with the root mean square difference, between each scanners result and against the matching ground truth. The cuboid phantom analysis showed a statistical difference for the thickness measurement in both scanners and in the length measurement for the optical scanner (p<0.05). However, no statistical difference was found between the scanner measurements. For the breast-like phantoms, the higher maximum distances were found in the infrared scans, but the mean distance between ground truth surface and the scans showed equivalent performance for both scanners. Our results suggest that the smartphone-based SLSS performance is sufficient to be used to create a complete three-dimensional model of the breast shape.

Description of purpose Contrast-enhanced spectral mammography can be used to guide needle biopsies. However, in vertical approach the compressed breast is deformed generating a so-called bump in the paddle aperture, which may interfere with the visibility of contrast-uptakes. Local thickness estimation would provide an enhanced image quality of the recombined image, increasing the visibility of the contrast-uptakes to be targeted during the biopsy procedure. In this work we propose a method to estimate the shape of the breast bump in biopsy vertical approach. Materials and Methods Our method consists on two steps: first, we compute a raw thickness which does not take into account the presence of contrast-uptakes; second, we use a physical model to separate the sparse iodine texture from the breast shape. This physical model is composed by a sum of Fourier components, describing the main shape of the bump, a series of low-order polynomials, describing the main compressed thickness, paddle tilt and deflection, and non-linear components describing the translation and rotation of the paddle aperture. A 3D object mimicking a bump was fabricated to test the pertinence of our shape model. Also, clinical images of 21 patients which followed CESM-guided biopsy were visually assessed. Results Comparison between raw and final estimated thickness of our 3D test object shows an error standard deviation of 0.37 mm similar to the noise standard deviation equals to 0.32 mm. The visual assessment of clinical cases showed that the thickness correction removes the superimposed low-frequency pattern due to non-uniform thickness of the bump, improving the identification of the lesion to be targeted. Conclusion The proposed method for thickness estimation is adapted to CESM-guided biopsies in vertical approach and it improves the identification of the contrast-uptakes that need to be targeted during the procedure.

The aim of the study is to assess the feasibility of a homemade phantom for image quality evaluation in Contrast Enhanced Spectral Mammography (CESM). The phantom was composed by a PMMA slab with holes of different diameters (10, 5 and 2.5 mm) and thicknesses (5, 4, 3 and 2 mm) filled with diluted iodine contrast medium, resulting in concentrations of 1.9, 1.5, 1.1 and 0.7 mg/cm² (±0.2 mg/cm2 ), similar to the clinical concentrations. Furthermore, we added tissue-equivalent slabs with anatomical background and we simulated 3 different configurations equivalent to 32, 60 and 90 mm breast thicknesses. Image acquisitions were performed on a Hologic 3Dimensions mammography system using AEC clinical parameters. The acquisitions included a low energy exposure followed by an high energy one, and the resulting processed images were a subtraction of the 2 acquired images. For each configuration, the CNR on the low, high and subtracted images were calculated. The results showed that CNR values measured on the processed subtracted images were much higher respect to the CNR measured on the “for processing” low and high energy images. Furthermore, as expected, an increase in CNR for increasing iodine concentration was verified on the processed images, but not always on raw images that contained anatomical background. Preliminary results showed that the phantom is suitable for image quality evaluation in CESM but further studies with different acquisition parameters and on different mammography systems are necessary to assess the repeatability and the consistency of the measurements.

Objects created by 3D printers are increasingly used in various medical applications. Today, affordable 3D printers, using Fused Deposition Modeling are widely available. In this project, a commercially available 3D printer was used to replicate a conventional radiographic contrast detail phantom. Printing materials were selected by comparing their x-ray attenuation properties. Two replicas were printed using polylactic acid, with different filling patterns. The printed phantoms were imaged by a clinical mammography system, using automatic exposure control. Phantom images were visually and quantitively compared to images of the corresponding conventional contrast detail phantom. Visual scoring of the contrast detail elements was performed by a medical physics student. Contrast-to-noise ratio (CNR) was calculated for each phantom element. The diameter and thickness of the smallest visible phantom object were 0.44 mm and 0.09 mm, respectively, for both filling patterns. For the conventional phantom, the diameter and thickness of the smallest visible object were 0.31 mm and 0.09 mm. Visual inspection of printed phantoms revealed some linear artefacts. These artefacts were however not visible on mammographic projections. Quantitively, average CNR of printed phantom objects followed the same trend with an increase of average CNR with increasing disk height. However, there is a limitation of detail objects with disk diameters below 1.25 mm, caused by the available nozzle size. Based upon the encouraging results, future work will explore the use of different materials and smaller nozzle diameters.

False positive recall is a major issue in breast cancer screening and the introduction of artificial intelligence (AI) might affect which women who are unnecessarily recalled. We have investigated how an AI system works on false positive recalls at screening and compared with radiologist findings. Two-view digital mammography (DM) examinations from 656 recalled women (136 with screening detected cancer), were analysed with a commercial AI system. The AI findings were matched with the areas on the images causing the recalls. The agreement was studied both at the examination level and for individual findings. Scores were compared between true positive and false positive recalls. ROC analysis was used to study the AI-system’s ability to distinguish between true and false positive recalls. It was also studied how the AI system performed on cases where there were discordant readings. AI identified the same areas as radiologists in 80% of the cases recalled on DM. For true positives both the proportion of matching areas and AI scores were higher than for false positive recalls. The AI system also had a relatively large AUC (0.83) for differentiating between false positive recalls and cancers. Further, the AI system identified most of the findings leading to recall in cases where only one of the readers had marked the case for discussion. There is a relatively large agreement between the AI system and radiologists. The AI system scores the false positives lower than true positives. AI complements a single reader in a way similar to a second reader.

Our lab has built a next-generation tomosynthesis (NGT) system utilizing scanning motions with more degrees of freedom than clinical digital breast tomosynthesis systems. We are working toward designing scanning motions that are customized around the locations of suspicious findings. The first step in this direction is to demonstrate that these findings can be detected with a single projection image, which can guide the remainder of the scan. This paper develops an automated method to identify findings that are prone to be masked. Perlin-noise phantoms and synthetic lesions were used to simulate masked cancers. NGT projections of phantoms were simulated using ray-tracing software. The risk of masking cancers was mapped using the ground-truth labels of phantoms. The phantom labels were used to denote regions of low and high risk of masking suspicious findings. A U-Net model was trained for multiclass segmentation of phantom images. Model performance was quantified with a receiver operating characteristic (ROC) curve using area under the curve (AUC). The ROC operating point was defined to be the point closest to the upper left corner of ROC space. The output predictions showed an accurate segmentation of tissue predominantly adipose (mean AUC of 0.93). The predictions also indicate regions of suspicious findings; for the highest risk class, mean AUC was 0.89, with a true positive rate of 0.80 and a true negative rate of 0.83 at the operating point. In summary, this paper demonstrates with virtual phantoms that a single projection can indeed be used to identify suspicious findings.

Digital mammography (DM) and digital breast tomosynthesis, the gold standards for breast cancer screening, requires correct breast positioning to ensure accuracy. Improper positioning can result in missed cancers, or can lead to additional imaging. We propose an automated deep learning (DL) segmentation approach to perform multi-class identification of regions of interest (ROI) commonly used for identification of poor positioning in mediolateral oblique (MLO) breast views. We hypothesize that by leveraging the capabilities of DL through the use of the well-founded U-Net model architecture, multi-class DL-based segmentation approaches can accurately identify air, parenchyma, pectoralis, and nipple locations within MLO images. In this study, we employed model hyperparameter searches to determine optimal model parameters for our proposed DL architecture, including the optimal loss function configuration; our best model achieved an average Sørensen-Dice coefficient of 0.919 ± 0.061 on the held-out test set. We identified high levels of localization performance in the nipple ROI. We believe our proposed segmentation model can be a foundational step in further mammogram analysis, such as for breast positioning and localized image processing tools.

As deep learning models are increasingly applied in medical diagnostic assistance systems, this raises questions about ones ability to understand and interpret its decision-making process. In this work, using breast lesions from the Optimam Mammography Image Database (OMI-DB), we have explored whether deep learned features have similar predictive information as classical texture features. We trained a deep learning model for mass lesion classification and used Gradient-weighted Class Activation Mapping to produce a representation of deep learned features. Additional, classical texture features (e.g. energy) were extracted. Subsequently, we used the earth mover’s distance to investigate similarities between deep learned and texture features. The comparison identified that texture features such as mean, entropy and auto-correlation showed a strong similarity with the deep learned features and provided an indication of what the deep learning models might have used as information for its classification.

In breast cancer detection, change in findings throughout time is one of the major biomarkers for the presence of malignancy. Several studies have established the value of comparing mammograms with the ones from previous examinations. Some of them have shown that such comparison decreases the recall rate and increases the biopsy yield of cancer but does not increase the cancer detection rate. This evidence brought us to do the hypotheses that, as for human radiologists, adding temporal context information could be beneficial also for artificial intelligence (AI) systems for breast cancer detection thus improving their specificity which today represents the major limitation for an autonomous use of such AI systems. In this study we carry out a comparison between an AI system for breast cancer detection and an update version of the same system able to integrate the temporal context information.

The aim of the DeepLook project, funded by INFN (Italy), is to implement a deep learning architecture for Computed Aided Detection (CAD), based on neural networks developed with deep learning methods, for the automatic detection and classification of breast lesions in DBT images. A preliminary step (started 2 years ago and still ongoing) was the creation of a dataset of annotated images. This dataset includes images acquired with different clinical DBT units and different acquisition geometries, on several hundred patients, containing a variety of possible breast lesions and normal cases of absence of lesions. This will make the diagnostic capacity of the CAD system particularly extensive in various clinical situations and on a significant sample of patients, so allowing the network to diagnose various types of lesions (at the level of the single tomosynthesis slices) and capable of operate on commercial DBT systems, also available from different vendors, as found in breast diagnosis departments. The developed CAD and first result of the indication of the slice containing the suspected mass will be presented.

Early detection of breast cancer in mammography screening via deep-learning based computer-aided detection systems shows promising potential in improving the curability and mortality rates of breast cancer. However, many clinical centres are restricted in the amount and heterogeneity of available data to train such models to (i) achieve promising performance and to (ii) generalise well across acquisition protocols and domains. As sharing data between centres is restricted due to patient privacy concerns, we propose a potential solution: sharing trained generative models between centres as substitute for real patient data. In this work, we use three well known mammography datasets to simulate three different centres, where one centre receives the trained generator of Generative Adversarial Networks (GANs) from the two remaining centres in order to augment the size and heterogeneity of its training dataset. We evaluate the utility of this approach on mammography patch classification on the test set of the GAN-receiving centre using two different classification models, (a) a convolutional neural network and (b) a transformer neural network. Our experiments demonstrate that shared GANs notably increase the performance of both transformer and convolutional classification models and highlight this approach as a viable alternative to inter-centre data sharing.

In recent years, the amount of images to be read has increased due to the higher resolution of diagnostic imaging devices, and the burden on doctors has also increased. To solve this problem, the improvement of CAD (computer-aided diagnosis) performance has been studied. In this study, we developed an AI-based system for discriminating benign and malignant breast cancer tumors using transfer learning, one of the deep learning methods of AI, and analyzed what factors are necessary to improve the diagnostic accuracy of the system. Classification of benign and malignant diseases using diagnostic images showed an accuracy of 90%, which was equivalent to physician's discrimination, but the accuracy for medical checkup images was low at 85%, and image comparison revealed that this was due to noise and low contrast. We analyzed that these improvements are necessary for the construction of a more accurate CAD system for medical checkup images.

The effectiveness of pharmacotherapy has been improved through the development of drugs that incorporate the knowledge of the molecular biology of breast cancer. Therefore, neoadjuvant chemotherapy (NAC) is actively administered to patients who wish to undergo breast-conserving surgery. During NAC, some patients have a pathological complete response (pCR). This study aims to develop a method for predicting patients with pCR during NAC. This creates new value for preoperative imaging. Breast ultrasound images were collected from 43 patients with breast cancer who received NAC at the Kumamoto University Hospital. The tumor area on the breast ultrasound image was manually marked. From the marked tumor regions, 379 radiomics features related to size, shape, density, and texture were measured. We employed the least absolute shrinkage and selection operator to select the useful radiomic features. Linear discriminant analysis (LDA) with eight selected radiomic features was used to distinguish between pCR and non-pCR. Leave-one-out was used for training and testing LDA. The sensitivity, specificity, and AUC were 89.5 % (17/19), 83.3% (19/24), and 0.920, respectively. Because the LDA is the simplest classifier, the phenotype of the lesion in the breast ultrasound image may contain information that predicts the therapeutic effect. Our proposed method could provide a new value for preoperative imaging.

Many studies have assessed breast density in clinical practice. However, calculation of breast density requires segmentation of the mammary gland region, and deep learning has only recently been applied. Thus, the robustness of the deep learning model for different image processing types has not yet been reported. We investigated the accuracy of segmentation of the U-net for mammograms made with variousimage processing types. We used 478 mediolateral oblique view mammograms. The mammograms were divided into 390 training images and 88 testing images. The ground truth of the mammary gland region made by mammary experts was used for the training and testing datasets. Four types of image processing (Types 1–4) were applied to the testing images to compare breast density in the segmented mammary gland regions with that of ground truths. The shape agreement between ground truth and the segmented mammary gland region by U-net of Types 1–4 was assessed using the Dice coefficient, and the equivalence or compatibility of breast density with ground truth was assessed by Bland-Altman analysis. The mean Dice coefficients between the ground truth and U-net were 0.952, 0.948, 0.948, and 0.947 for Types 1, 2, 3, and 4, respectively. By Bland-Altman analysis, the equivalence of breast density between ground truth and U-net was confirmed for Types 1 and 2, and compatibility was confirmed for Types 3 and 4. We concluded that the robustness of the U-net for segmenting the mammary gland region was confirmed for different image processing types.

Deep learning (DL) models can be trained on contrast-enhanced mammography (CEM) images to detect and classify lesions in the breast. As they often put more emphasis on the masses enhanced in the recombined image, they can fail in recognizing microcalcification clusters since these are hardly enhanced and are mainly visible in the (processed) lowenergy image. Therefore, we developed a method to create synthetic data with simulated microcalcification clusters to be used for data augmentation and explainability studies when training DL models. At first 3-dimensional voxel models of simulated microcalcification clusters based on descriptors of the shape and structure were constructed. In a set of 500 simulated microcalcification clusters the range of the size and of the number of microcalcifications per cluster followed the distribution of real clusters. The insertion of these clusters in real images of non-delineated CEM cases was evaluated by radiologists. The realism score was acceptable for single view applications. Radiologists could more easily categorize synthetic clusters into benign versus malignant than real clusters. In a second phase of the work, the role of synthetic data for training and/or explaining DL models was explored. A Mask R-CNN model was trained with synthetic CEM images containing microcalcification clusters. After a training run of 100 epochs the model was found to overfit on a training set of 192 images. In an evaluation with multiple test sets, it was found that this high level of sensitivity was due to the model being capable of recognizing the image rather than the cluster. Synthetic data could be applied for more tests, such as the impact of particular features in both background and lesion models.

Breast cancer is one of the deadliest diseases. It is affecting majority of women world wide. Computer Aided Diagnosis (CAD) systems can be used to help radiologists in order to examine the initial symptoms. One of the early symptoms is micro-calcifications. Detection of abnormalities is an essential part of treatment in the right direction. Along with detection of abnormalities, the classification of micro-calcification has a vital importance. Timely detection and classification of micro-calcification as malignant or benign can save a lot of women. We have used region based convolutional neural networks and obtained 92.7% mean average precision at training time while at testing time mAP is 89.2%.

The paper presents a framework for the detection of mass-like lesions in 3D digital breast tomosynthesis. It consists of several steps, including pre and post-processing, and a main detection block based on a Faster RCNN deep learning network. In addition to the framework, the paper describes different training steps to achieve better performance, including transfer learning using both mammographic and DBT data. The presented approach obtained third place in the recent DBT Lesion detection Challenge, DBTex, being the top approach without using an ensemble based method.

Triple-negative is one of the most aggressive type of breast cancer for which is also difficult to find an effective treatment. An early diagnosis and a fast and specific treatment are shown to be key aspects for a better prognosis. Current diagnosis of these cases are based on performing a biopsy. This study proposes a non-invasive medical imaging predication method, based on a deep learning architecture, to automatically classify triple-negative tumors in DCE-MRI images. Results are evaluated on an extensive public dataset for different normalizations, data augmentations, learning rates and batch sizes, reaching a state-of-the-art AUC of 0.68.

Success of breast cancer treatment is subject to various factors, including cancer stage and cancer grade. The best treatment is selected based on the characteristic of cancer. It is desirable to predict the cancer characteristics and prognostic factors accurately and promptly by diagnostic imaging. The purpose of the study is to investigate the use of multimodality diagnostic images in predicting breast cancer subtypes to assist diagnosis and treatment planning. In this study, we classify lesions into molecular subtypes and simultaneously predict histological grades and invasiveness of the cancers by mammography and breast ultrasound images. Models with different architectures including single input and multi-input layers with single head and multiple head models are compared. The results indicate that use of multimodality images is more predictive than using single modalities. The automatic subtype classification using multimodality images may support a prompt treatment planning and proper patient care.

This study aimed at conducting a review of the prior mammograms of screen-detected breast cancers, found on full-field digital mammograms based on independent double reading with arbitration. The prior mammograms of 607 women diagnosed with breast cancer during routine breast cancer screening were categorized into “Missed”, “Prior Vis”, and “Prior Invis” . The prior mammograms of “Missed” and “Prior Vis” cases showed actionable and non-actionable visible cancer signs, respectively. The “Prior Invis” cases had no overt cancer signs on the prior mammograms. The percentage of cases classified as “Missed”, “Prior Vis”, and “Prior Invis” categories were 25.5%, 21.7%, 52.7%, respectively. The proportion of high-density cases showed no significant differences among the three categories (p-values<0.05). The breakdown of cases into “Missed”, “Prior Vis”, and “Prior Invis” categories did not differ between invasive (488) and in-situ (119) cases. In the invasive category, the progesterone (p-value=0.015) and estrogen (p-value=0.007) positivity and the median ki-67 score (p-value=0.006) differed significantly among the categories with the “Prior Invis” cases exhibiting the highest percentage of hormone receptors negativity. In the invasive cases, the percentage of cancers graded as 3 (i.e., more aggressive) were significantly more in the “Prior Invis” category compared to both “Missed” and “Prior Vis” categories (both p-values<0.05). The status of receptors and breast cancer grade for the in-situ cases did not differ significantly among the three categories. Prior images categorization can predict the aggressiveness of breast cancer. Techniques to better interrogate prior images as shown elsewhere may yield important patient outcomes.

Screening programs for the early detection of breast cancer have significantly reduced mortality in women. The limitations of these programmes are primarily due to the use of 2D techniques and the high number of mammograms to be read by radiologists. Artificial Intelligence (AI) systems may lead to new tools to help radiologists read mammograms and classify the examination based on the malignancy of the detected lesions. Several factors related to breast characteristics (thickness and density), technical factors of image acquisition, X-ray system performance and image processing algorithms can influence the outcome of a mammogram and thus also the detection capability of an AI system. The aim of this work is to analyze the robustness of an AI system for breast cancer detection and its dependence on breast characteristics and technical factors. For this purpose, mammograms from a population-based screening program were scored with the AI system. The AUC (area under the ROC curve) index generated from the scoring ROC curve was 0.92 (CI(95%) = 0.89 - 0.95), demonstrating the robust performance of the AI system. Moreover, the statistical analysis performed showed that the AUC index was independent of breast characteristics, the type of mammographic system and most of the technical parameters considered, demonstrating the effectiveness of the AI system.

Objectives: To study the effect on radiology trainees’ observer performance through the availability of prior screening mammograms as part of seven unique education test sets. Methods: Australian radiology trainees (n=150) completed 469 readings of seven educational test sets (each set with 60 cases, 40 normal and 20 cancer cases). The percentage of cases with a prior screening mammogram was 68.7%. Mammographic density (MD) evaluated via BIRADS was spread across the test sets, with 40.5% having 25-50% glandular tissue (BIRADS “B”), 37.4% of cases having 50-75% or “C”, 12.6% have a >75% MD and 9.5% having the lowest MD rating “A”. Trainees were asked to score the cases on a scale of 1 (normal), 2 (benign), 3 (equivocal findings), 4 (suspicious finding) and 5 (highly suggestive malignancy). Mann-Whitney U was used to compare the specificity and sensitivity of radiology trainees among cases with and without prior images. Results: Radiology trainees had significantly higher sensitivity across all MD levels when prior images were not available (A-B, P=0.006; C-D, P=0.027). Specificity was also significantly higher for cases of high (C-D) MD without prior images compared with priors available by trainees who read less than 20 cases per week (P=0.008). Conclusions: In a simulated environment, radiology trainees achieved better results in cases without prior images, especially for those who read less than 20 cases per week. The utility of prior case inclusion when providing education and training in reading screening mammograms needs to be revisited, especially for women with high MD.

Purpose: To identify mammographic image quality indicators (IQI) predictive of interval breast cancers (IC) as opposed to screen-detected cancers (SDC). Methods: Eligible cases for the study were raw, routine recall, screening exams acquired at two UK sites between 2010- 2018, from the OPTIMAM database. Women were matched 3:1 (SDC, n=965 versus IC, n=326), by age (nearest), screening site, breast density grade, Xray system vendor, and compression paddle. Images of the affected breast for prior (IC only) or incident (SDC only) exams were processed using automated software to obtain volumetric breast density (VBD) and IQI metrics related to compression and breast positioning. Compression pressure (CP) was categorised into tertiles or low/target/high (<7/7-15/<15 kPa) groups. Univariate and logistic regression analyses were used to identify significant predictors of IC versus SDC. Results: Compared to SDC, IC had lower median CP (7.9 versus 8.6 kPa, p<0.05). Multivariate analysis found only CP to be significantly associated with the risk of IC versus SDC, with odds ratios (OR) and 95% confidence intervals of 0.93 (0.89-0.97) per unit CP. Compared to low CP, target CP was significantly associated with a lower IC versus SDC risk at the breast level [OR=0.73 (0.56-0.95)] and for mediolateral oblique views [OR=0.77 (0.59-0.99)]. Comparing the third and first tertile, CP was significantly associated with lower risk of IC versus SDC [0.64 (0.47-0.87)], with very similar results when analysed per view. Conclusions: CP was found to be a significant predictor of IC versus SDC, with higher CP being associated with a lower risk of IC.

Tumor growth rate estimations can provide useful information about tumor progression and aggressiveness. Understanding the breast cancer progression and aggressiveness could aid with personalized screening/follow-up, treatment options, and prognosis. This paper reports a preliminary estimation of the tumor volume doubling time (TVDT) for cancers detected during the Malmö Breast Tomosynthesis Screening Trial (MBTST). The trial included 14 848 women in whom 139 cancers were detected. Out of those, 101 spiculated or circumscribed masses, had prior images available, making them suitable for tumor growth evaluation. In the preliminary analysis of images from 30 women, tumor size was measured in mammograms from MBTST and prior images. The analyzed cases were selected among women with visible tumors in two consecutive screening exams. The tumor size was measured in two orthogonal directions. The average of the two measurements was used in the analysis. The mean time and the corresponding standard deviation (SD) between the two consecutive mammograms were 744 ± 73 days. The mean TVDT and SD were 637 ± 428 days (range 159-2373 days). Future work will include the analysis of a larger number of women and a stratification of TVDT related to screening intervals.

Aim: This project is part of a long-term goal to apply radiomics-based risk prediction models designed for twodimensional (2D) digital mammography (DM) to three-dimensional (3D) digital breast tomosynthesis (DBT), using either the DBT projection views (PV) or the reconstructed planes. In this work, 2 fundamental aspects related to PVs were explored: (1) finding robust radiomic features for both DM and PV, and (2) selecting robust and informative radiomic features for both 2D and 3D modalities by requiring respectively invariance and noninvariance of these features across DBT projections. Methods: DM and PVs from combined DM and DBT acquisitions of phantom and patients were used in this study. Robust radiomic features in these images were identified by the intra-class correlation coefficient (ICC) between DM and the central PV for DBT. Then, projection invariant and noninvariant radiomic features of PVs for different projection angles were also characterized by ICC. Finally, selected projection invariant features of PVs were applied on a DM breast density classifier and their predictive power was compared to the results of DM. Results: A total of 70 out of 93 extracted radiomic features (75%) showed at least moderate reliability (ICC>0.5) between DM and the central PV. In addition, a decrease of feature reliability along increasing angular range was observed on both real and simulated datasets. With projection angle invariance as the feature selection method, overfitting of a DM density classifier was reduced. Conclusions: A large portion of radiomic features was robust between DM and the central PV without specific harmonization, suggesting that some parts of the radiomic features of DM can be applied to the DBT projection dataset. Additionally, 3D DBT could also benefit 2D DM through the projection angle variation test. Projectioninvariant features with better robustness could be selected for 2D DM which was preliminary validated by a density classification task, while projection non-invariant features which incorporate 3D information in the PVs may be suitable for 3D DBT.

Aim: To assess the difference in radiomic feature values between pairs of mammographic images used for processing(FOR PROC) and for presentation(FOR PRES) as well as the ability to determine the BIRADS density classification from these radiomic features with different classification models. Methods: A dataset of FOR PROC and FOR PRES image pairs annotated with labels for the BI-RADS classification done by a radiologist is used in this study. The differences in radiomic feature values between the image types are evaluated with the intraclass correlation coefficient(ICC). Additionally, the discriminative power of radiomic feature values regarding the BI-RADS score is evaluated with Logistic Regression, Random Forest and a 5-layer deep Neural Network. The results of these models are evaluated with a 5-fold crossvalidation. Results: The reliability of radiomic feature is generally low between pairs of FOR PROC and FOR PRES images for all radiomic feature groups. Furthermore, the simple models used to determine the ability to assign the BI-RADS density classification based on the radiomic feature values reached insufficient accuracy to be considered adequate. Conclusion: The study revealed low reliability between both image types. Furthermore radiomic features alone seem to be insufficient to determine the BI-RADS classification using simple models.

Purpose: To introduce a novel technique for pretraining deep neural networks on mammographic images, where the network learns to predict multiple metadata attributes and simultaneously to match images from the same patient and study. Further to demonstrate how this network can be used to produce explainable predictions. Methods: We trained a neural network on a dataset of 85,558 raw mammographic images and seven types of metadata, using a combination of supervised and self-supervised learning techniques. We evaluated the performance of our model on a dataset of 4,678 raw mammographic images using classification accuracy and correlation. We also designed an ablation study to demonstrate how the model can produce explainable predictions. Results: The model learned to predict all but one of the seven metadata fields with classification accuracy ranging from 78-99% on the validation dataset. The model was able to predict which images were from the same patient with over 93% accuracy on a balanced dataset. Using a simple X-ray system classifier built on top of the first model, representations learned on the initial X-ray system classification task showed by far the largest effect size on ablation, illustrating a method for producing explainable predictions. Conclusions: It is possible to train a neural network to predict several kinds of mammogram metadata simultaneously. The representations learned by the model for these tasks can be summed to produce an image representation that captures features unique to a patient and study. With such a model, ablation offers a promising method to enhance the explainability of deep learning predictions.

The aim of this retrospective case-cohort study was to perform additional validation of an artificial intelligence (AI)-driven breast cancer risk model in a racially diverse cohort of women undergoing screening. We included 176 breast cancer cases with non-actionable mammographic screening exams 3 months to 2 years prior to cancer diagnosis and a random sample of 4,963 controls from women with non-actionable mammographic screening exams and at least one-year of negative follow-up (Hospital University Pennsylvania, PA, USA; 9/1/2010-1/6/2015). A risk score for each woman was extracted from full-field digital mammography (FFDM) images via an AI risk prediction model, previously developed and validated in a Swedish screening cohort. The performance of the AI risk model was assessed via age-adjusted area under the ROC curve (AUC) for the entire cohort, as well as for the two largest racial subgroups (White and Black). The performance of the Gail 5-year risk model was also evaluated for comparison purposes. The AI risk model demonstrated an AUC for all women = 0.68 95% CIs [0.64, 0.72]; for White = 0.67 [0.61, 0.72]; for Black = 0.70 [0.65, 0.76]. The AI risk model significantly outperformed the Gail risk model for all women (AUC = 0.68 vs AUC = 0.55, p<0.01) and for Black women (AUC = 0.71 vs AUC = 0.48, p<0.01), but not for White women (AUC = 0.66 vs AUC = 0.61, p=0.38). Preliminary findings in an independent dataset suggest a promising performance of the AI risk prediction model in a racially diverse breast cancer screening cohort.

X-ray imaging results in inhomogeneous irradiation of the detector and distortion of structures in the periphery of the image; yet the spatial dependency of tomosynthesis image-quality metrics has not been extensively investigated. In this study, we use virtual clinical trials to quantify the spatial dependency of lesion detectability in our lab’s next-generation tomosynthesis (NGT) system. Two geometries were analyzed: a conventional geometry with mediolateral source motion, and a NGT geometry with T-shaped motion. Breast parenchymal texture was simulated using an open-source library with Perlin noise using 400 random seeds and three breast densities. Spherical mass lesions were inserted in the central slice of the phantoms using the voxel additive method. Image acquisition was simulated using in-house ray-tracing software and simple backprojection was performed using commercial reconstruction software. Lesion detectability with Channelized Hotelling Observers (CHOs) was analyzed using receiver operating characteristic curves to measure the detectability index (d') at 154 unique locations for the lesions. We also divided images into three non-overlapping regions (differing in terms of distance from the chest wall). At the 0.05 level of significance, there was a statistically significant difference between the geometries in terms of d' in one of the three regions, with the T geometry offering superior detectability. Examining all 154 lesion locations, the T geometry was found to offer lower spread (standard deviation) in d' values throughout the image area, and superior d' at 83 of 154 locations (53.9%). In summary, the T geometry enables superior lesion detection and mitigates anisotropies.

This work aims at evaluating the spatial resolution and noise in 3D images acquired with a clinical Computed Tomography scanner dedicated to the breast (BCT). The presampled modulation transfer function (MTF) and the noise power spectrum (NPS) are measured. In addition, the capability of the system in showing simulated lesions and microcalcification clusters was assessed via a phantom test. The impact of the selected reconstruction algorithm on MTF, NPS, and simulated lesion visibility was evaluated. The available algorithms are the Standard (Std) and Calcification (Calc) reconstructions, which use an isotropic reconstructed voxel edge of 0.273 mm and the high-resolution (HR) reconstruction algorithm that uses an isotropic reconstructed voxel edge of 0.190 mm. The spatial frequency (expressed in mm^-1 ) at which the MTF curve goes down to 10% (MTF_10%) was found to be 1.0 mm^-1 for the case of Std reconstruction in radial direction at the chest-wall; this value increases to 1.3 mm^-1 and 1.5 mm^-1 for the HR and Calc reconstructions, respectively. The distance from the isocenter did not impact the system spatial resolution. As expected, the improvement in the spatial resolution in the Calc and HR reconstruction algorithms is accompanied by an increase in the noise, especially at the higher frequencies, as shown in the 1D NPS. A phantom study showed that both simulated soft lesion with diameter of 1.8 mm and microcalcification cluster with grain diameter of 0.29 mm are visible, no matter what reconstruction algorithm is selected. Microcalcifications with diameter of 0.20 mm and 0.13 mm do not appear to be visible.

A method is proposed to equate the measured noise to a thickness of Aluminium by exposing a simple test object which includes an Aluminium step-wedge sandwiched between PMMA slabs. The scaling process turns the Image Noise expressed in mmAl into an absolute quantity which reflects exposure conditions. A second quantity, the dose independent Normalized Image Noise is defined. It is a characteristic value for every mammography unit/model and phantom setup and represents a measure of a system’s overall detection efficiency in clinical conditions. 8 mammography units of different manufacturers and detector technologies have been evaluated for PMMA thickness 3, 5 and 7 cm over a wide average glandular dose (AGD) range. The calculated Normalised Image Noise values were reproducible (uncertainty 3-5%) and coherent with known physical characteristics of the detector-grid combinations. Image Noise resulted sensitive to radiation spectra and scatter amount. Starting from threshold contrast detection with the CDMAM ver. 3.4 phantom, it was possible to identify Image Noiseacceptable/achievable thresholds which correspond to adequate image quality. Signal difference to noise (SDNR) analysis based on the proposed test object was in good agreement with SDNR evaluation according to the EUREF guideline (difference 1- 7%). Conversion coefficient, Image Noise and Normalised Image Noise could all be derived from a single exposure without having to determine the detector's response curve beforehand which is particularly advantageous for non-linear response. Given the sensitivity of Image Noise to radiation quality and dose, it is a suitable metric for optimization.

It is well-known that x-ray systems featuring indirect detectors are affected by noise spatial correlation. In the case of digital breast tomosynthesis (DBT), this phenomenon might affect the perception of small details in the image, such as microcalcifications. In this work, we propose the use of a deep convolutional neural network (CNN) to restore DBT projections degraded with correlated noise using the framework of a cycle generative adversarial network (cycle-GAN). To generate pairs of images for the training procedure, we used a virtual clinical trial (VCT) system. Two approaches were evaluated: in the first one, the network was trained to perform noise decorrelation by changing the frequency-dependency of the noise in the input image, but keeping the other characteristics. In the second approach, the network was trained to perform denoising and decorrelation, with the objective of generating an image with frequency-independent (white) noise and with characteristics equivalent to an acquisition with a radiation exposure four times greater than the input image. We tested the network with virtual and clinical images and we found that in both training approaches the model successfully corrected the power spectrum of the input images.

Digital breast tomosynthesis (DBT) is an important imaging modality for breast cancer screening. The morphology of breast masses and the shape of the microcalcifications are important factors to detect and determine the malignancy of breast cancer. Recently, convolutional neural networks (CNNs) have been used for denoising in medical imaging and have shown potential to improve the performance of radiologists. However, they can impose noise spatial correlation in the restoration process. Noise correlation can negatively impact radiologists’ performance, creating image signals that can resemble breast lesions. In this work, we propose a deep CNN that restores low-dose DBT projections by partially filtering out the noise, but imposes fidelity of the noise correlation between the original and restored images, avoiding artifacts that may resemble signs of breast cancer. The combination of a loss function that calculates the difference in the power spectra (PS) of the input and output images and another one that seeks image visual perception is proposed. We compared the performance of the proposed neural network with traditional denoising methods that do not consider the noise correlation in the restoration process and found superior results in terms of PS for our approach.

The angular range and number of projections are parameters that directly influence the image quality and the visibility of lesions in digital breast tomosynthesis (DBT). The medical field is taking advantage of the increasing performance of machine learning algorithms with the use of complex data-driven models, known as deep learning (DL) networks. The use of DL has also been highlighted in the tasks of video frame interpolation (VFI) for the synthesis of new images in order to increase the frame rate per second. In the present work, we use a residual refinement interpolation network (RRIN) to generate new synthetic DBT projections from pairs of real projections. We studied two different approaches: first, we increased the number of projections before reconstruction using the synthetic images, with the aim of improving the quality of the reconstructed slices without increasing the radiation dose to the patient. In the second, we investigated the effect of replacing existing projections with synthetic ones, with the objective of reducing the radiation dose and acquisition time. In the first approach, we used virtual phantoms to generate sets of DBT projections to train the network. We then evaluated the contrast-to-noise ratio (CNR) of simulated microcalcifications after reconstruction. The CNR was higher for all sets where supplementary images were added compared to those with only real images. In the second approach, we trained the network with clinical data and tested it with images acquired with a physical anthropomorphic breast phantom. Both the projections and the slices showed good similarity with the real ones, suggesting that the use of VFI networks to generate DBT projections is promising. However, further studies should be carried out to assess the feasibility of this approach.