Faculty Bibliography

Deep neural networks (DNNs) show promise in image-based medical diagnosis, but cannot be fully trusted since they can fail for reasons unrelated to underlying pathology. Humans are less likely to make such superficial mistakes, since they use features that are grounded on medical science. It is therefore important to know whether DNNs use different features than humans. Towards this end, we propose a framework for comparing human and machine perception in medical diagnosis. We frame the comparison in terms of perturbation robustness, and mitigate Simpson's paradox by performing a subgroup analysis. The framework is demonstrated with a case study in breast cancer screening, where we separately analyze microcalcifications and soft tissue lesions. While it is inconclusive whether humans and DNNs use different features to detect microcalcifications, we find that for soft tissue lesions, DNNs rely on high frequency components ignored by radiologists. Moreover, these features are located outside of the region of the images found most suspicious by radiologists. This difference between humans and machines was only visible through subgroup analysis, which highlights the importance of incorporating medical domain knowledge into the comparison.

Launching an academic career in breast imaging presents both challenges and opportunities for the newly graduated trainee. A strategic plan aligned with one's personal strengths and interests facilitates career success and professional satisfaction. Academic departments offer multiple tracks to accommodate diverse faculty goals. The specific requirements of various tracks vary across institutions. The clinician-educator track typically encourages a focus on medical education and educational scholarship. The clinician-investigator or clinician-scholar track supports original research and grant-funded clinical trials. Finally, the clinical and clinician-administrator tracks allow for emphasis on clinical program development and leadership. As definitions of scholarship broaden, many opportunities are accessible to demonstrate excellence in the traditional areas of clinical practice, education, and research, as well as the broader fields of leadership and administration. Departmental and national society resources that advance knowledge in one's chosen area of interest are available and should be explored. Mentorship and sponsorship can provide valuable insight into identifying such resources and devising a plan for sustainable career success and work-life integration.

The management of intraductal papilloma (IDP) diagnosed on core needle biopsy (CNB) is controversial due to the variable upgrade rates to breast carcinoma (BC) on subsequent surgical excision reported in the literature. The purpose of our study was to investigate the upgrade rate of IDP diagnosed on CNB to BC in subsequent surgical excision and the impact of clinical, pathologic and radiologic variables. This is a retrospective cohort of all women who had a diagnosis of IDP on a CNB between 2005 and 2018 in a tertiary academic center with subsequent surgical excision. Upgrade was defined as ductal carcinoma in situ (DCIS) and invasive carcinoma on surgical excision. Statistical analyses included Pearson's chi-square, Wilcoxon rank-sum and logistic regression. A total of 216 women with IDP in a CNB were included. Nineteen patients (8.8%) upgraded to BC in the overall cohort, including 14 DCIS and 5 invasive carcinomas. An upgrade rate of 27% was found in atypical IDP (14 of 51 cases), while only 3% of pure IDP upgraded to BC (5 of 165 cases). Older age (>53 years) at time of biopsy (OR=1.05, 95%CI 1.01-1.09, p=0.027) and concomitant atypical ductal hyperplasia (ADH) (OR=9.69, 95%CI 3.37-27.81, p<0.0001) were significantly associated with upgrade. Our results support surgical excision of IDP on CNB when associated with ADH or diagnosed in women older than 53 years of age. The low surgical upgrade rate of 3% for pure IDP on CNB in younger women should be part of the management discussion.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 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. (i) 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. (ii) 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. (iii) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. (iv) 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/breastcancerclassifier.

BACKGROUND:Mammographic breast density (MBD) is an independent risk factor for breast cancer. Information regarding the relationship of MBD and breast cancer biology in women with ductal carcinoma in situ (DCIS) is currently lacking. This study aimed to examine the clinicopathologic characteristics of DCIS in women stratified by MBD. METHODS:A retrospective review was performed to identify women with pure DCIS who underwent preoperative mammography between 2010 and 2018. Clinicopathologic and demographic data were collected. For the purpose of analysis, MBD was categorized as "non-dense" (Breast Imaging-Reporting and Data System [BI-RADS] density categories A and B) or "dense" (BI-RADS C and D) according to its identification in radiology reports. RESULTS:Of 3227 patients with a breast cancer diagnosis enrolled in the institutional Breast Cancer Database during the study period, 658 (20%) had pure DCIS. Of these 658 patients, 42% had non-dense breasts, and 58% had dense breasts. Most lesions were non-palpable (92%) and detected by mammography (84%). Patients with dense breasts were more likely to be younger at the time of diagnosis (p < 0.001), premenopausal (p < 0.001), and Asian (p = 0.018), and to have higher-grade disease (p = 0.006; Table 2). Family history, BRCA status, parity, mammogram frequency, palpability, method of presentation, lesion size, hormone receptor status, comedo histology, and recurrence did not differ significantly between the two groups (Table 1). The median follow-up period was 7.1 years. CONCLUSION/CONCLUSIONS:Women with pure DCIS and higher MBD are more likely to be younger at the time of diagnosis, premenopausal, and Asian, and to present with higher-grade disease. Further research on the relationship of age, MBD, and tumor biology in DCIS is warranted.

BACKGROUND:Oncologic outcomes with nipple-sparing mastectomy (NSM) continue to be established. We examine oncologic trends, outcomes, and risk factors, including tumor-to-nipple distance (TND), in therapeutic NSMs. METHODS:Demographics, outcomes, and overall trends for all NSMs undertaken for a therapeutic indication from 2006 to 2017 were analyzed. Oncologic outcomes were investigated with specific focus on recurrence and associated factors, including TND. RESULTS:A total of 496 therapeutic NSMs were performed with average follow-up time of 48.25 months. The most common tumor types were invasive carcinoma (52.4%) and ductal carcinoma in situ (50.4%). Sentinel lymph node sampling was performed in 79.8% of NSMs; 4.1% had positive frozen sentinel lymph node biopsies while 15.7% had positive nodal status on permanent pathologic examination. The most common pathologic cancer stage was stage IA (42.5%) followed by Stage 0 (31.3%).Per NSM, the rate of local recurrence was 1.6% (N=8); the rate of regional recurrence was 0.6% (N=3). In all, 171 NSMs had magnetic resonance imaging available to assess tumor-to-nipple distance (TND). NSMs with TND ≤1 centimeter (25.0% versus 2.4%, p=0.0031/p=0.1129) and ≤2 centimeters (8.7% versus 2.0%; p=0.0218/p=0.1345) trended to higher rates of locoregional recurrence. In univariate analysis, TND ≤1 centimeter was the only significant risk factor for recurrence (OR=13.5833, p=0.0385). No factors were significant in regression analysis. CONCLUSIONS:In this group of early stage and in situ breast carcinoma, therapeutic NSM appears oncologically safe with a locoregional recurrence rate of 2.0%. Tumor-to-nipple distances of ≤1 centimeter and ≤2 centimeters trended to higher rates of recurrence.

PURPOSE: The study sought to determine screening mammography recommendations that radiologists in Canada promote to average-risk patients and family or friends, and do or would do for themselves. METHODS: An online survey was delivered from February 19, 2014, to July 11, 2014. Data included radiologists' recommendations for mammography and their personal screening habits based on gender. The 3 radiologists' cohorts were women >/=40 years of age, women <40 years of age, and men. The distribution of responses for each question was summarized, and proportions for the entire group and individual cohorts were computed. RESULTS: Of 402 surveys collected, 97% (299 of 309) radiologists recommended screening every 1-2 years, 62% (192 of 309) starting >/=40 years of age and 2% (5 of 309) recommended screening every 2-3 years for women 50-74 years of age. Recommendations were similar for family and friends: 96% (294 of 305) recommended screening every 1-2 years, 66% (202 of 305) recommended screening every 1-2 years for women >/=40 years of age, and 2% (5 of 305) recommended screening every 2-3 years. For women radiologists >/=40 years of age, 76% (48 of 63) underwent screening every 1-2 years and started at 40 years of age, 76% (16 of 21) female radiologists <40 years of age would undergo screening >/=40 years of age, 100% every 1-2 years, and 90% (151 of 167) male radiologists would undergo screening every 1-2 years, with 71% (120 of 169) beginning at 40 years of age. CONCLUSION: The majority of Canadian radiologists recommend screening mammography every 1-2 years for average-risk women >/=40 years of age, whether they are patients or family and friends.

Mentoring is an extremely important component of academic medicine, including radiology, yet it is not specifically emphasized in radiology training, and many academic radiology departments in the United States, Canada, and elsewhere do not have formal mentoring programs for medical students, residents, fellows, or junior faculty. The purpose of this article is to overview the current status of mentorship in radiology, to discuss the importance of mentorship at multiple levels and its potential benefits in particular, as well as how to conduct a successful mentor-mentee relationship. The literature on mentorship in radiology and in academic medicine in general is reviewed.

OBJECTIVE: The purpose of this study is to determine which patient- and tumor-related and clinical variables influence dedicated breast surgeons' and general surgeons' referrals for preoperative breast MRI for patients with newly diagnosed breast cancer. MATERIALS AND METHODS: Surgeons who perform breast surgery responded to a survey from June 16, 2014, through August 11, 2014. Participants self-identified as breast or general surgeons and provided professional practice details. They used Likert scores (range, 1-7 with increasing likelihood to order MRI) to weigh numerous patient- and tumor-related and clinical variables. Mean likelihood scores were calculated and compared using a linear mixed model. A p