Improving lesion detection in mammograms by leveraging a Cycle-GAN-based lesion remover.

BACKGROUND: The wide heterogeneity in the appearance of breast lesions and normal breast structures can confuse computerized detection algorithms. Our purpose was therefore to develop a Lesion Highlighter (LH) that can improve the performance of computer-aided detection algorithms for detecting breast cancer on screening mammograms. METHODS: We hypothesized that a Cycle-GAN based Lesion Remover (LR) could act as an LH, which can improve the performance of lesion detection algorithms. We used 10,310 screening mammograms from 4,832 women that included 4,942 recalled lesions (BI-RADS 0) and 5,368 normal results (BI-RADS 1). We divided the dataset into Train:Validate:Test folds with the ratios of 0.64:0.16:0.2. We segmented image patches (400 × 400 pixels) from either lesions marked by MQSA radiologists or normal tissue in mammograms. We trained a Cycle-GAN to develop two GANs, where each GAN transferred the style of one image to another. We refer to the GAN transferring the style of a lesion to normal breast tissue as the LR. We then highlighted the lesion by color-fusing the mammogram after applying the LR to its original. Using ResNet18, DenseNet201, EfficientNetV2, and Vision Transformer as backbone architectures, we trained three deep networks for each architecture, one trained on lesion highlighted mammograms (Highlighted), another trained on the original mammograms (Baseline), and Highlighted and Baseline combined (Combined). We conducted ROC analysis for the three versions of each deep network on the test set. RESULTS: The Combined version of all networks achieved AUCs ranging from 0.963 to 0.974 for identifying the image with a recalled lesion from a normal breast tissue image, which was statistically improved (p-value < 0.001) over their Baseline versions with AUCs that ranged from 0.914 to 0.967. CONCLUSIONS: Our results showed that a Cycle-GAN based LR is effective for enhancing lesion conspicuity and this can improve the performance of a detection algorithm.

Standalone AI for Breast Cancer Detection at Screening Digital Mammography and Digital Breast Tomosynthesis: A Systematic Review and Meta-Analysis.

Background There is considerable interest in the potential use of artificial intelligence (AI) systems in mammographic screening. However, it is essential to critically evaluate the performance of AI before it can become a modality used for independent mammographic interpretation. Purpose To evaluate the reported standalone performances of AI for interpretation of digital mammography and digital breast tomosynthesis (DBT). Materials and Methods A systematic search was conducted in PubMed, Google Scholar, Embase (Ovid), and Web of Science databases for studies published from January 2017 to June 2022. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) values were reviewed. Study quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 and Comparative (QUADAS-2 and QUADAS-C, respectively). A random effects meta-analysis and meta-regression analysis were performed for overall studies and for different study types (reader studies vs historic cohort studies) and imaging techniques (digital mammography vs DBT). Results In total, 16 studies that include 1 108 328 examinations in 497 091 women were analyzed (six reader studies, seven historic cohort studies on digital mammography, and four studies on DBT). Pooled AUCs were significantly higher for standalone AI than radiologists in the six reader studies on digital mammography (0.87 vs 0.81, P = .002), but not for historic cohort studies (0.89 vs 0.96, P = .152). Four studies on DBT showed significantly higher AUCs in AI compared with radiologists (0.90 vs 0.79, P < .001). Higher sensitivity and lower specificity were seen for standalone AI compared with radiologists. Conclusion Standalone AI for screening digital mammography performed as well as or better than radiologists. Compared with digital mammography, there is an insufficient number of studies to assess the performance of AI systems in the interpretation of DBT screening examinations. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Scaranelo in this issue.

Breast MRI contrast enhancement kinetics of normal parenchyma correlate with presence of breast cancer.

BACKGROUND: We investigated dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) contrast enhancement kinetic variables quantified from normal breast parenchyma for association with presence of breast cancer, in a case-control study. METHODS: Under a Health Insurance Portability and Accountability Act compliant and Institutional Review Board-approved protocol, DCE-MRI scans of the contralateral breasts of 51 patients with cancer and 51 controls (matched by age and year of MRI) with biopsy-proven benign lesions were retrospectively analyzed. Applying fully automated computer algorithms on pre-contrast and multiple post-contrast MR sequences, two contrast enhancement kinetic variables, wash-in slope and signal enhancement ratio, were quantified from normal parenchyma of the contralateral breasts of both patients with cancer and controls. Conditional logistic regression was employed to assess association between these two measures and presence of breast cancer, with adjustment for other imaging factors including mammographic breast density and MRI background parenchymal enhancement (BPE). The area under the receiver operating characteristic curve (AUC) was used to assess the ability of the kinetic measures to distinguish patients with cancer from controls. RESULTS: When both kinetic measures were included in conditional logistic regression analysis, the odds ratio for breast cancer was 1.7 (95 % CI 1.1, 2.8; p = 0.017) for wash-in slope variance and 3.5 (95 % CI 1.2, 9.9; p = 0.019) for signal enhancement ratio volume, respectively. These odds ratios were similar on respective univariate analysis, and remained significant after adjustment for menopausal status, family history, and mammographic density. While percent BPE was associated with an odds ratio of 3.1 (95 % CI 1.2, 7.9; p = 0.018), in multivariable analysis of the three measures, percent BPE was non-significant (p = 0.897) and the two kinetics measures remained significant. For the differentiation of patients with cancer and controls, the unadjusted AUC was 0.71 using a combination of the two measures, which significantly (p = 0.005) outperformed either measure alone (AUC = 0.65 for wash-in slope variance and 0.63 for signal enhancement ratio volume). CONCLUSIONS: Kinetic measures of wash-in slope and signal enhancement ratio quantified from normal parenchyma in DCE-MRI are jointly associated with presence of breast cancer, even after adjustment for mammographic density and BPE.

Organizational Breast Cancer Data Mart: A Solution for Assessing Outcomes of Imaging and Treatment.

PURPOSE: In the United States, a comprehensive national breast cancer registry (CR) does not exist. Thus, care and coverage decisions are based on data from population subsets, other countries, or models. We report a prototype real-world research data mart to assess mortality, morbidity, and costs for breast cancer diagnosis and treatment. METHODS: With institutional review board approval and Health Insurance Portability and Accountability Act (HIPPA) compliance, a multidisciplinary clinical and research data warehouse (RDW) expert group curated demographic, risk, imaging, pathology, treatment, and outcome data from the electronic health records (EHR), radiology (RIS), and CR for patients having breast imaging and/or a diagnosis of breast cancer in our institution from January 1, 2004, to December 31, 2020. Domains were defined by prebuilt views to extract data denormalized according to requirements from the existing RDW using an export, transform, load pattern. Data dictionaries were included. Structured query language was used for data cleaning. RESULTS: Five-hundred eighty-nine elements (EHR 311, RIS 211, and CR 67) were mapped to 27 domains; all, except one containing CR elements, had cancer and noncancer cohort views, resulting in a total of 53 views (average 12 elements/view; range, 4-67). EHR and RIS queries returned 497,218 patients with 2,967,364 imaging examinations and associated visit details. Cancer biology, treatment, and outcome details for 15,619 breast cancer cases were imported from the CR of our primary breast care facility for this prototype mart. CONCLUSION: Institutional real-world data marts enable comprehensive understanding of care outcomes within an organization. As clinical data sources become increasingly structured, such marts may be an important source for future interinstitution analysis and potentially an opportunity to create robust real-world results that could be used to support evidence-based national policy and care decisions for breast cancer.

Stereoscopic digital mammography: improved specificity and reduced rate of recall in a prospective clinical trial.

PURPOSE: To compare stereoscopic digital mammography (DM) with standard DM for the rate of patient recall and the detection of cancer in a screening population at elevated risk for breast cancer. MATERIALS AND METHODS: Starting in September 2004 and ending in December 2007, this prospective HIPAA-compliant, institutional review board-approved screening trial, with written informed consent, recruited female patients at elevated risk for breast cancer (eg, personal history of breast cancer or breast cancer in a close relative). A total of 1298 examinations from 779 patients (mean age, 58.6 years; range, 32-91 years) comprised the analyzable data set. A paired study design was used, with each enrolled patient serving as her own control. Patients underwent both DM and stereoscopic DM examinations in a single visit, findings of which were interpreted independently by two experienced radiologists, each using a Breast Imaging Reporting and Data System (BI-RADS) assessment (BI-RADS category 0, 1, or 2). All patients determined to have one or more findings with either or both modalities were recalled for standard diagnostic evaluation. The results of 1-year follow-up or biopsy were used to determine case truth. RESULTS: Compared with DM, stereoscopic DM showed significantly higher specificity (91.2% [1167 of 1279] vs 87.8% [1123 of 1279]; P = .0024) and accuracy (90.9% [1180 of 1298] vs 87.4% [1135 of 1298]; P = .0023) for detection of cancer. Sensitivity for detection of cancer was not significantly different for stereoscopic DM (68.4% [13 of 19]) compared with DM (63.2% [12 of 19], P .99). The recall rate for stereoscopic DM was 9.6% (125 of 1298) and that for DM was 12.9% (168 of 1298) (P = .0018). CONCLUSION: Compared with DM, stereoscopic DM significantly improved specificity for detection of cancer, while maintaining comparable sensitivity. The recall rate was significantly reduced with stereoscopic DM compared with DM. SUPPLEMENTAL MATERIAL: http://radiology.rsna.org/lookup/suppl/doi:10.1148/radiol.12120382/-/DC1.

Impact of GAN artifacts for simulating mammograms on identifying mammographically occult cancer.

Purpose: Generative adversarial networks (GANs) can synthesize various feasible-looking images. We showed that a GAN, specifically a conditional GAN (CGAN), can simulate breast mammograms with normal, healthy appearances and can help detect mammographically-occult (MO) cancer. However, similar to other GANs, CGANs can suffer from various artifacts, e.g., checkerboard artifacts, that may impact the quality of the final synthesized image, as well as the performance of detecting MO cancer. We explored the types of GAN artifacts that exist in mammogram simulations and their effect on MO cancer detection. Approach: We first trained a CGAN using digital mammograms (FFDMs) of 1366 women with normal/healthy breasts. Then, we tested the trained CGAN on an independent MO cancer dataset with 333 women with dense breasts (97 MO cancers). We trained a convolutional neural network (CNN) on the MO cancer dataset, in which real and simulated mammograms were fused, to identify women with MO cancer. Then, a radiologist who was independent of the development of the CGAN algorithms evaluated the entire MO cancer dataset to identify and annotate artifacts in the simulated mammograms. Results: We found four artifact types, including checkerboard, breast boundary, nipple-areola complex, and black spots around calcification artifacts, with an overall incidence rate over 69% (the individual incident rate ranged from 9% to 53%) from both normal and MO cancer samples. We then evaluated their potential impact on MO cancer detection. Even though various artifacts existed in the simulated mammogram, we found that it still provided complementary information for MO cancer detection when it was combined with the real mammograms. Conclusions: We found that artifacts were pervasive in the CGAN-simulated mammograms. However, they did not negatively affect our MO cancer detection algorithm; the simulated mammograms still provided complementary information for MO cancer detection when combined with real mammograms.

Use of Artificial Intelligence for Digital Breast Tomosynthesis Screening: A Preliminary Real-world Experience.

OBJECTIVE: The purpose of this study is to assess the "real-world" impact of an artificial intelligence (AI) tool designed to detect breast cancer in digital breast tomosynthesis (DBT) screening exams following 12 months of utilization in a subspecialized academic breast center. METHODS: Following IRB approval, mammography audit reports, as specified in the BI-RADS atlas, were retrospectively generated for five radiologists reading at three locations during a 12-month time frame. One location had the AI tool (iCAD ProFound AI v2.0), and the other two locations did not. The co-primary endpoints were cancer detection rate (CDR) and abnormal interpretation rate (AIR). Secondary endpoints included positive predictive values (PPVs) for cancer among screenings with abnormal interpretations (PPV1) and for biopsies performed (PPV3). Odds ratios (OR) with two-sided 95% confidence intervals (CIs) summarized the impact of AI across radiologists using generalized estimating equations. RESULTS: Nonsignificant differences were observed in CDR, AIR, and PPVs. The CDR was 7.3 with AI and 5.9 without AI (OR 1.3, 95% CI: 0.9-1.7). The AIR was 11.7% with AI and 11.8% without AI (OR 1.0, 95% CI: 0.8-1.3). The PPV1 was 6.2% with AI and 5.0% without AI (OR 1.3, 95% CI: 0.97-1.7). The PPV3 was 33.3% with AI and 32.0% without AI (OR 1.1, 95% CI: 0.8-1.5). CONCLUSION: Although we are unable to show statistically significant changes in CDR and AIR outcomes in the two groups, the results are consistent with prior reader studies. There is a nonsignificant trend toward improvement in CDR with AI, without significant increases in AIR.

Retrieval boosted computer-aided diagnosis of clustered microcalcifications for breast cancer.

PURPOSE: The authors propose an image-retrieval based approach for case-adaptive classifier design in computer-aided diagnosis (CADx). The conventional approach in CADx is to first train a pattern-classifier based on a set of existing training samples and then apply this classifier to subsequent new cases. The purpose of this work is to improve the classification accuracy of a CADx classifier by making use of a set of known cases retrieved from a reference library that are similar to the case under consideration. METHODS: In the proposed approach, the authors will first apply image-retrieval to obtain a set of lesion images from a library of known cases that have similar image features to a case being diagnosed (i.e., query). These retrieved cases are then used to optimize a pattern-classifier toward boosting its classification accuracy on the query case. The basic idea is to put more emphasis on those cases that are similar to the query. The proposed approach is demonstrated first using a linear classifier and then extended to a nonlinear classifier induced by kernel principal component analysis. RESULTS: The proposed retrieval-driven approach was tested on a library of mammogram images from 1006 cases (646 benign and 360 malignant) obtained from multiple institutions and was demonstrated to yield significant improvement in classification performance. Measured by the area under the receiver operating characteristic curve (AUC), the case-adaptive approach could boost the classification performance of a linear classifier from AUC = 0.7415 to AUC = 0.7807; similar improvement was also obtained for a nonlinear classifier, with AUC boosted from 0.7527 to 0.7838. CONCLUSIONS: Use of additional cases from a reference library that have similar image features can improve the classification accuracy of a CADx classifier on a query case. It can even outperform retraining the classifier with all the cases from the entire reference library. This implies that cases with similar image features are more relevant in defining the local decision boundary of the CADx classifier around the query.

A statistically defined anthropomorphic software breast phantom.

PURPOSE: Digital anthropomorphic breast phantoms have emerged in the past decade because of recent advances in 3D breast x-ray imaging techniques. Computer phantoms in the literature have incorporated power-law noise to represent glandular tissue and branching structures to represent linear components such as ducts. When power-law noise is added to those phantoms in one piece, the simulated fibroglandular tissue is distributed randomly throughout the breast, resulting in dense tissue placement that may not be observed in a real breast. The authors describe a method for enhancing an existing digital anthropomorphic breast phantom by adding binarized power-law noise to a limited area of the breast. METHODS: Phantoms with (0.5 mm)(3) voxel size were generated using software developed by Bakic et al. Between 0% and 40% of adipose compartments in each phantom were replaced with binarized power-law noise (ß = 3.0) ranging from 0.1 to 0.6 volumetric glandular fraction. The phantoms were compressed to 7.5 cm thickness, then blurred using a 3 × 3 boxcar kernel and up-sampled to (0.1 mm)(3) voxel size using trilinear interpolation. Following interpolation, the phantoms were adjusted for volumetric glandular fraction using global thresholding. Monoenergetic phantom projections were created, including quantum noise and simulated detector blur. Texture was quantified in the simulated projections using power-spectrum analysis to estimate the power-law exponent ß from 25.6 × 25.6 mm(2) regions of interest. RESULTS: Phantoms were generated with total volumetric glandular fraction ranging from 3% to 24%. Values for ß (averaged per projection view) were found to be between 2.67 and 3.73. Thus, the range of textures of the simulated breasts covers the textures observed in clinical images. CONCLUSIONS: Using these new techniques, digital anthropomorphic breast phantoms can be generated with a variety of glandular fractions and patterns. ß values for this new phantom are comparable with published values for breast tissue in x-ray projection modalities. The combination of conspicuous linear structures and binarized power-law noise added to a limited area of the phantom qualitatively improves its realism.

A comparison study of image features between FFDM and film mammogram images.

PURPOSE: This work is to provide a direct, quantitative comparison of image features measured by film and full-field digital mammography (FFDM). The purpose is to investigate whether there is any systematic difference between film and FFDM in terms of quantitative image features and their influence on the performance of a computer-aided diagnosis (CAD) system. METHODS: The authors make use of a set of matched film-FFDM image pairs acquired from cadaver breast specimens with simulated microcalcifications consisting of bone and teeth fragments using both a GE digital mammography system and a screen-film system. To quantify the image features, the authors consider a set of 12 textural features of lesion regions and six image features of individual microcalcifications (MCs). The authors first conduct a direct comparison on these quantitative features extracted from film and FFDM images. The authors then study the performance of a CAD classifier for discriminating between MCs and false positives (FPs) when the classifier is trained on images of different types (film, FFDM, or both). RESULTS: For all the features considered, the quantitative results show a high degree of correlation between features extracted from film and FFDM, with the correlation coefficients ranging from 0.7326 to 0.9602 for the different features. Based on a Fisher sign rank test, there was no significant difference observed between the features extracted from film and those from FFDM. For both MC detection and discrimination of FPs from MCs, FFDM had a slight but statistically significant advantage in performance; however, when the classifiers were trained on different types of images (acquired with FFDM or SFM) for discriminating MCs from FPs, there was little difference. CONCLUSIONS: The results indicate good agreement between film and FFDM in quantitative image features. While FFDM images provide better detection performance in MCs, FFDM and film images may be interchangeable for the purposes of training CAD algorithms, and a single CAD algorithm may be applied to either type of images.

Clinically missed cancer: how effectively can radiologists use computer-aided detection?

OBJECTIVE: The purpose of this study was to determine the effectiveness with which radiologists can use computer-aided detection (CADe) to detect cancer missed at screening. MATERIALS AND METHODS: An observer study was performed to measure the ability of radiologists to detect breast cancer on mammograms with and without CADe. The images in the study were from 300 analog mammographic examinations. In 234 cases the mammograms were read clinically as normal and free of cancer for at least 2 subsequent years. In the other 66 cases, cancers were missed clinically. In 256 cases, current and previous mammograms were available. Eight radiologists read the dataset and recorded a BI-RADS assessment, the location of the lesion, and their level of confidence that the patient should be recalled for diagnostic workup for each suspicious lesion. Jackknife alternative free-response receiver operating characteristic analysis was used. RESULTS: The jackknife alternative free-response receiver operating characteristic figure of merit was 0.641 without aid and 0.659 with aid (p = 0.06; 95% CI, -0.001 to 0.036). The sensitivity increased 9.9% (95% CI, 3.4-19%) and the callback rate 12.1% (95% CI, 7.3-20%) with CADe. Both increases were statistically significant (p < 0.001). Radiologists on average ignored 71% of correct computer prompts. CONCLUSION: Use of CADe can increase radiologist sensitivity 10% with a comparable increase in recall rate. There is potential for CADe to have a bigger clinical impact because radiologists failed to recognize a correct computer prompt in 71% of missed cancer cases [corrected].

Assessing the stand-alone sensitivity of computer-aided detection with cancer cases from the Digital Mammographic Imaging Screening Trial.

OBJECTIVE: The purpose of this study was to assess the sensitivities and false-detection rates of two computer-aided detection (CADe) systems when applied to digital or film-screen mammograms in detecting the known breast cancer cases from the Digital Mammographic Imaging Screening Trial (DMIST) breast cancer screening population. MATERIALS AND METHODS: Available film-screen and digital mammograms of 161 breast cancer cases from DMIST were analyzed by two CADe systems, iCAD Second-Look and R2 ImageChecker. Three experienced breast-imaging radiologists reviewed the CADe marks generated for each available cancer case, recording the number and locations of CADe marks and whether each CADe mark location corresponded with the known location of the cancer. RESULTS: For the 161 cancer cases included in this study, the sensitivities of the DMIST reader without CAD were 0.43 (69/161, 95% CI 0.35-0.51) for digital and 0.41 (66/161, 0.33-0.49) for film-screen mammography. The sensitivities of iCAD were 0.74 (119/161, 0.66-0.81) for digital and 0.69 (111/161, 0.61-0.76) for film-screen mammography, both significantly higher than the DMIST study sensitivities (p < 0.0001 for both). The average number of false CADe marks per case of iCAD was 2.57 (SD, 1.92) for digital and 3.06(1.72) for film-screen mammography. The sensitivity of R2 was 0.74 (119/161, 0.66-0.81) for digital, and 0.60 (97/161, 0.52-0.68) for film-screen mammography, both significantly higher than the DMIST study sensitivities (p < 0.0001 for both). The average number of false CADe marks per case of R2 was 2.07 (1.57) for digital and 1.52 (1.45) for film-screen mammography. CONCLUSION: Our results suggest the use of CADe in interpretation of digital and film-screen mammograms could lead to improvements in cancer detection.

Identifying Women With Mammographically- Occult Breast Cancer Leveraging GAN-Simulated Mammograms.

Our objective is to show the feasibility of using simulated mammograms to detect mammographically-occult (MO) cancer in women with dense breasts and a normal screening mammogram who could be triaged for additional screening with magnetic resonance imaging (MRI) or ultrasound. We developed a Conditional Generative Adversarial Network (CGAN) to simulate a mammogram with normal appearance using the opposite mammogram as the condition. We used a Convolutional Neural Network (CNN) trained on Radon Cumulative Distribution Transform (RCDT) processed mammograms to detect MO cancer. For training CGAN, we used screening mammograms of 1366 women. For MO cancer detection, we used screening mammograms of 333 women (97 MO cancer) with dense breasts. We simulated the right mammogram for normal controls and the cancer side for MO cancer cases. We created two RCDT images, one from a real mammogram pair and another from a real-simulated mammogram pair. We finetuned a VGG16 on resulting RCDT images to classify the women with MO cancer. We compared the classification performance of the CNN trained on fused RCDT images, CNNFused to that of trained only on real RCDT images, CNNReal, and to that of trained only on simulated RCDT images, CNNSimulated. The test AUC for CNNFused was 0.77 with a 95% confidence interval (95CI) of [0.71, 0.83], which was statistically better (p-value < 0.02) than the CNNReal AUC of 0.70 with a 95CI of [0.64, 0.77] and CNNSimulated AUC of 0.68 with a 95CI of [0.62, 0.75]. It showed that CGAN simulated mammograms can help MO cancer detection.

Developing breast lesion detection algorithms for digital breast tomosynthesis: Leveraging false positive findings.

BACKGROUND: Due to the complex nature of digital breast tomosynthesis (DBT) in imaging techniques, reading times are longer than 2D mammograms. A robust computer-aided diagnosis system in DBT could help radiologists reduce their workload and reading times. PURPOSE: The purpose of this study was to develop algorithms for detecting biopsy-proven breast lesions on DBT using multi-depth level convolutional models and leveraging non-biopsied samples. As biopsied positive samples in a lesion dataset are limited, we hypothesized that false positive (FP) findings by detection algorithms from non-biopsied benign lesions could improve detection algorithms by using them as data augmentation. APPROACH: We first extracted 2D slices from DBT volumes with biopsy-proven breast lesions (cancer and benign), with non-biopsied benign lesions (actionable), and for controls. Then, to provide lesion continuity along the z-direction, we combined a lesion slice with its immediate adjacent slices to synthesize 2.5-dimensional (2.5D) images of the lesion by assigning them into R, G, and B color channels. We used 224 biopsy-proven lesions from 39 cancer and 62 benign patients from a DBTex challenge dataset of 1000 scans. We included the 2.5D images of immediate neighboring slices from the lesion's center to increase the number of training samples. For lesion detection, we used the YOLOv5 algorithm as our base network. We trained a baseline algorithm (medium-depth level) using biopsied samples to detect actionable FPs in non-biopsied images. Afterward, we fine-tuned the baseline model on the augmented image set (actionable FPs added). For lesion inferencing, we processed the DBT volume slice-by-slice to estimate bounding boxes in each slice, and then combined them by connecting bounding boxes along the depth via volumetric morphological closing. We trained an additional model (large) with deeper-depth levels by repeating the above process. Finally, we developed an ensemble algorithm by combining the medium and large detection models. We used the free-response operating characteristic curve to evaluate our algorithms. We reported mean sensitivity per FPs per DBT volume only for biopsied views and sensitivity at 2-false positives per image (2FPI) for all views. However, due to the limited accessibility to the truth of the challenge validation and test datasets, we used sensitivity at 2FPI for statistical evaluation. RESULTS: For the DBTex independent validation set, the medium baseline model achieved a mean sensitivity of 0.627 FPs per DBT volume, and a sensitivity of 0.640 at 2FPI. After adding actionable FP lesions, the model had an improved 2FPI of 0.769 over the baseline (p-value = 0.013). Our ensemble algorithm with multi-depth levels (medium + large) achieved a mean sensitivity of 0.815 FPs per DBT volume and an improved sensitivity at 2FPI of 0.80 over the baseline (p-value < 0.001) on the validation set. Finally, our ensemble model achieved a mean sensitivity of 0.786 FPs per DBT volume and a sensitivity of 0.743 at 2FPI on the DBTex independent test set. CONCLUSIONS: Our results show that actionable FP findings hold useful information for lesion detection algorithms, and our ensemble detection model with multi-depth levels improves lesion detection performance.

Using the Medical Audit to Improve Practice Performance.

Feedback to physicians on their clinical performance is critical to continuous learning and maintenance of skills as well as maintaining patient safety. However, it is fraught with challenges around both implementation and acceptance. Additionally, rewarding of performance improvement is not often done, putting into question the efficacy of the process. Physician audit and feedback have been studied extensively and shown to be beneficial in many fields of medicine. Documenting physician performance and sharing individual and group data have been positively linked to changing physician behavior, ultimately leading to improved patient outcomes. Although casual review of one's own performance is often the easiest approach, it is frequently over- or underestimated by self-evaluation. Objective measures are therefore important to provide concrete data on which physicians can act. A fundamental question remains in mammography: Is reporting the information to the physician and accreditation bodies enough, or should there be consequences for the radiologist and/or facility if there is outlier behavior?

Fifty years of SPIE Medical Imaging proceedings papers.

Purpose: To commemorate the 50th anniversary of the first SPIE Medical Imaging meeting, we highlight some of the important publications published in the conference proceedings. Approach: We determined the top cited and downloaded papers. We also asked members of the editorial board of the Journal of Medical Imaging to select their favorite papers. Results: There was very little overlap between the three methods of highlighting papers. The downloads were mostly recent papers, whereas the favorite papers were mostly older papers. Conclusions: The three different methods combined provide an overview of the highlights of the papers published in the SPIE Medical Imaging conference proceedings over the last 50 years.

Contrast enhancement of hepatic hemangiomas on multiphase MDCT: Can we diagnose hepatic hemangiomas by comparing enhancement with blood pool?

OBJECTIVE: The purpose of this article is to determine whether enhancement of nodular foci within hemangiomas is homogeneous and matches blood vessels at different phases on contrast-enhanced MDCT. MATERIALS AND METHODS: Multiphase (unenhanced, arterial, portal venous, and delayed phases) MDCT images of 58 hemangiomas were reviewed by two radiologists. Nodular-enhancing foci within hemangiomas were evaluated for enhancement pattern and were subjectively compared with enhancement of the aorta, inferior vena cava, hepatic vein, and portal vein for each contrast-enhanced phase. Both readers measured CT attenuation of enhancing nodules and vessels at each phase, and enhancement of nodules and vessels was compared. RESULTS: Qualitative analysis showed heterogeneously enhancing nodules in 79.3% and 65.5% of hemangiomas in the arterial phase and in 74.1% and 53.4% of hemangiomas in the portal venous phase, according to readers 1 and 2, respectively. In the arterial phase, 3.8% and 12.3% of nodules showed enhancement similar to that in the aorta. In the portal venous phase, 15.4% and 21.7%, 16.8% and 18.2%, 14.1% and 23.8%, and 19.5% and 25.9% of nodules were scored with enhancement similar to that in the aorta, inferior vena cava, hepatic vein, and portal vein by readers 1 and 2, respectively. Differences between attenuation of nodules and all vessels in the arterial, portal venous, and delayed phases were statistically significant. Statistically significant differences were also noted between attenuation among blood vessels in the arterial and portal venous phases but not in the delayed phase. CONCLUSION: Attenuation of enhancing foci within hemangiomas does not match vessel density qualitatively or quantitatively. No common blood pool density exists in the arterial or portal venous phase. Although persistent enhancement without washout is a useful CT criterion, specific criteria to match the blood pool cannot be used to confirm a diagnosis of hemangioma.

Cross-organ, cross-modality transfer learning: feasibility study for segmentation and classification.

We conducted two analyses by comparing the transferability of a traditionally transfer-learned CNN (TL) to that of a CNN fine-tuned with an unrelated set of medical images (mammograms in this study) first and then fine-tuned a second time using TL, which we call the cross-organ, cross-modality transfer learned (XTL) network, on 1) multiple sclerosis (MS) segmentation of brain magnetic resonance (MR) images and 2) tumor malignancy classification of multi-parametric prostate MR images. We used 2133 screening mammograms and two public challenge datasets (longitudinal MS lesion segmentation and ProstateX) as intermediate and target datasets for XTL, respectively. We used two CNN architectures as basis networks for each analysis and fine-tuned it to match the target image types (volumetric) and tasks (segmentation and classification). We evaluated the XTL networks against the traditional TL networks using Dice coefficient and AUC as figure of merits for each analysis, respectively. For the segmentation test, XTL networks outperformed TL networks in terms of Dice coefficient (Dice coefficients of 0.72 vs [0.70 - 0.71] with p-value < 0.0001 in differences). For the classification test, XTL networks (AUCs = 0.77 - 0.80) outperformed TL networks (AUC = 0.73 - 0.75). The difference in the AUCs (AUCdiff = 0.045 - 0.047) was statistically significant (p-value < 0.03). We showed XTL using mammograms improves the network performance compared to traditional TL, despite the difference in image characteristics (x-ray vs. MRI and 2D vs. 3D) and imaging tasks (classification vs. segmentation for one of the tasks).