An Efficient Deep Neural Network to Classify Large 3D Images With Small Objects.

3D imaging enables accurate diagnosis by providing spatial information about organ anatomy. However, using 3D images to train AI models is computationally challenging because they consist of 10x or 100x more pixels than their 2D counterparts. To be trained with high-resolution 3D images, convolutional neural networks resort to downsampling them or projecting them to 2D. We propose an effective alternative, a neural network that enables efficient classification of full-resolution 3D medical images. Compared to off-the-shelf convolutional neural networks, our network, 3D Globally-Aware Multiple Instance Classifier (3D-GMIC), uses 77.98%-90.05% less GPU memory and 91.23%-96.02% less computation. While it is trained only with image-level labels, without segmentation labels, it explains its predictions by providing pixel-level saliency maps. On a dataset collected at NYU Langone Health, including 85,526 patients with full-field 2D mammography (FFDM), synthetic 2D mammography, and 3D mammography, 3D-GMIC achieves an AUC of 0.831 (95% CI: 0.769-0.887) in classifying breasts with malignant findings using 3D mammography. This is comparable to the performance of GMIC on FFDM (0.816, 95% CI: 0.737-0.878) and synthetic 2D (0.826, 95% CI: 0.754-0.884), which demonstrates that 3D-GMIC successfully classified large 3D images despite focusing computation on a smaller percentage of its input compared to GMIC. Therefore, 3D-GMIC identifies and utilizes extremely small regions of interest from 3D images consisting of hundreds of millions of pixels, dramatically reducing associated computational challenges. 3D-GMIC generalizes well to BCS-DBT, an external dataset from Duke University Hospital, achieving an AUC of 0.848 (95% CI: 0.798-0.896).

An Objective Structured Clinical Examination for Medical Student Radiology Clerkships: Reproducibility Study.

BACKGROUND: Objective structured clinical examinations (OSCEs) are a useful method to evaluate medical students' performance in the clerkship years. OSCEs are designed to assess skills and knowledge in a standardized clinical setting and through use of a preset standard grading sheet, so that clinical knowledge can be evaluated at a high level and in a reproducible way. OBJECTIVE: This study aimed to present our OSCE assessment tool designed specifically for radiology clerkship medical students, which we called the objective structured radiology examination (OSRE), with the intent to advance the assessment of clerkship medical students by providing an objective, structured, reproducible, and low-cost method to evaluate medical students' radiology knowledge and the reproducibility of this assessment tool. METHODS: We designed 9 different OSRE cases for radiology clerkship classes with participating third- and fourth-year medical students. Each examination comprises 1 to 3 images, a clinical scenario, and structured questions, along with a standardized scoring sheet that allows for an objective and low-cost assessment. Each medical student completed 3 of 9 random examination cases during their rotation. To evaluate for reproducibility of our scoring sheet assessment tool, we used 5 examiners to grade the same students. Reproducibility for each case and consistency for each grader were assessed with a two-way mixed effects intraclass correlation coefficient (ICC). An ICC below 0.4 was deemed poor to fair, an ICC of 0.41 to 0.60 was moderate, an ICC of 0.6 to 0.8 was substantial, and an ICC greater than 0.8 was almost perfect. We also assessed the correlation of scores and the students' clinical experience with a linear regression model and compared mean grades between third- and fourth-year students. RESULTS: A total of 181 students (156 third- and 25 fourth-year students) were included in the study for a full academic year. Moreover, 6 of 9 cases demonstrated average ICCs more than 0.6 (substantial correlation), and the average ICCs ranged from 0.36 to 0.80 (P<.001 for all the cases). The average ICC for each grader was more than 0.60 (substantial correlation). The average grade among the third-year students was 11.9 (SD 4.9), compared with 12.8 (SD 5) among the fourth-year students (P=.005). There was no correlation between clinical experience and OSRE grade (-0.02; P=.48), adjusting for the medical school year. CONCLUSIONS: Our OSRE is a reproducible assessment tool with most of our OSRE cases showing substantial correlation, except for 3 cases. No expertise in radiology is needed to grade these examinations using our scoring sheet. There was no correlation between scores and the clinical experience of the medical students tested.

Breast Imaging in Transgender Patients: What the Radiologist Should Know.

Transgender is the umbrella term for individuals whose gender identity and/or gender expression differs from their assigned sex at birth. With the rise in patients undergoing gender-affirming hormone therapy and gender-affirming surgery, it is increasingly important for radiologists to be aware of breast imaging considerations for this population. While diagnostic imaging protocols for transgender individuals are generally similar to those for cisgender women, screening guidelines are more variable. Currently, several professional and institutional guidelines have been created to address breast cancer screening in the transgender population, specifically screening mammography in transfeminine individuals who undergo hormone therapy. This article defines appropriate terminology with respect to the transgender population, reviews evidence for breast cancer risk and screening in transgender individuals, considers diagnostic breast imaging approaches, and discusses special considerations and challenges with regard to health care access and public education for these individuals. ©RSNA, 2019.

Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening.

We present a deep convolutional neural network for breast cancer screening exam classification, trained, and evaluated on over 200000 exams (over 1000000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. 1) Our network's novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. 2) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. 3) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. 4) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, the model's design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breast_cancer_classifier.

Ductal Carcinoma In Situ: The Whole Truth.

OBJECTIVE: Ductal carcinoma in situ (DCIS) is a noninvasive malignant breast disease traditionally described as a precursor lesion to invasive breast cancer. With screening mammography, DCIS now accounts for approximately 20% of newly diagnosed cancer cases. DCIS is not well understood because of its heterogeneous nature. CONCLUSION: Studies have aimed to assess prognostic factors to characterize its risk of invasive potential; however, there still remains a lack of uniformity in workup and treatment. We summarize current knowledge of DCIS and the ongoing controversies.

Citation Impact of Collaboration in Radiology Research.

PURPOSE: Team science involving multidisciplinary and multi-institutional collaboration is increasingly recognized as a means of strengthening the quality of scientific research. The aim of this study was to assess associations between various forms of collaboration and the citation impact of published radiology research. METHODS: In 2010, 876 original research articles published in Academic Radiology, the American Journal of Roentgenology, JACR, and Radiology were identified with at least one radiology-affiliated author. All articles were manually reviewed to extract features related to all authors' disciplines and institutions. Citations to these articles through September 2016 were extracted from Thomson Reuters Web of Science. RESULTS: Subsequent journal article citation counts were significantly higher (P < .05) for original research articles with at least seven versus six or fewer authors (26.2 ± 30.8 versus 20.3 ± 23.1, respectively), with authors from multiple countries versus from a single country (32.3 ± 39.2 versus 22.0 ± 25.0, respectively), with rather than without a nonuniversity collaborator (28.7 ± 38.6 versus 22.4 ± 24.9, respectively), and with rather than without a nonclinical collaborator (26.5 ± 33.1 versus 21.9 ± 24.4, respectively). On multivariate regression analysis, the strongest independent predictors of the number of citations were authors from multiple countries (ß = 9.14, P = .002), a nonuniversity collaborator (ß = 4.80, P = .082), and at least seven authors (ß = 4.11, P = .038). CONCLUSIONS: With respect to subsequent journal article citations, various forms of collaboration are associated with greater scholarly impact of published radiology research. To enhance the relevance of their research, radiology investigators are encouraged to pursue collaboration across traditional disciplinary, institutional, and geographic boundaries.

FDG PET/CT in Pancreatic and Hepatobiliary Carcinomas: Value to Patient Management and Patient Outcomes.

Fludeoxyglucose F 18 ((18)F-FDG) PET/CT has not been shown to offer additional benefit in the initial diagnosis of pancreatic cancer, but studies show benefit of (18)F-FDG PET/CT in initial staging and patient prognosis. There is evidence for (18)F-FDG PET and (18)F-FDG PET/CT in staging and prognosis of cholangiocarcinoma and gallbladder cancer. (18)F-FDG PET/CT has shown promise in staging liver malignancies by detecting extrahepatic metastasis. There is evidence supporting the ability of PET/CT in predicting prognosis in patients with hepatocellular carcinoma. Evidence is evolving for the role of (18)F-FDGPET/CT in predicting prognosis and survival in patients with colorectal liver metastasis.