Faculty Bibliography

This dataset, Collab-CXR, provides a unique resource to study human-AI collaboration in chest X-ray interpretation. We present experimentally generated data from 227 professional radiologists who assessed 324 historical cases under varying information conditions: with and without AI assistance, and with and without clinical history. Using a custom-designed interface, we collected probabilistic assessments for 104 thoracic pathologies using a comprehensive hierarchical reporting structure. This dataset is the largest known comparison of human-AI collaborative performance to either AI or humans alone in radiology, offering assessments across an extensive range of pathologies with rich metadata on radiologist characteristics and decision-making processes. Multiple experimental designs enable both within-subject and between-subject analyses. Researchers can leverage this dataset to investigate how radiologists incorporate AI assistance, factors influencing collaborative effectiveness, and impacts on diagnostic accuracy, speed, and confidence across different cases and pathologies. By enabling rigorous study of human-AI integration in clinical workflows, this dataset can inform AI tool development, implementation strategies, and ultimately improve patient care through optimized collaboration in medical imaging.

INTRODUCTION/UNASSIGNED:Interventions and surgical procedures are common for nonmalignant lung lesions detected on lung cancer screening (LCS). Inadvertent surgical resection of benign nodules with a clinical suspicion of lung cancer can occur, can be associated with complications, and adds to the cost of screening. The objective of this study is to assess the characteristics of surgically resected benign nodules detected on LCS computed tomography which were presumed to be lung cancers. METHODS/UNASSIGNED:This retrospective study included 4798 patients who underwent LCS between June 2014 and January 2021. The benign lung nodules, surgically resected with a presumed cancer diagnosis, were identified from the LCS registry. Patient demographics, imaging characteristics, and pathologic diagnoses of benign nodules were analyzed. RESULTS/UNASSIGNED:Of the 4798 patients who underwent LCS, 148 (3.1%) underwent surgical resection of a lung nodule, and of those who had a resection, 19 of 148 (12.8%) had a benign diagnosis (median age = 64 y, range: 56-77 y; F = 12 of 19, 63.2%; M = seven of 19, 36.8%). The median nodule size was 10 mm (range: 6-31 mm). Most nodules were solid (15 of 19, 78.9%), located in the upper lobes (11 of 19; 57.9%), and were peripheral (17 of 19, 89.5%). Most nodules (13 of 17; 76.5%) had interval growth, and four of 17 (23.5%) had increased fluorodeoxyglucose uptake. Of the 19 patients, 17 (89.5%) underwent sublobar resection (16 wedge resection and one segmentectomy), whereas two central nodules (10.5%) had lobectomies. Pathologies identified included focal areas of fibrosis or scarring (n = 8), necrotizing granulomatous inflammation (n = 3), other nonspecific inflammatory focus (n = 3), benign tumors (n = 3), reactive lymphoid hyperplasia (n = 1), and organizing pneumonia (n = 1). CONCLUSIONS/UNASSIGNED:Surgical resections of benign nodules that were presumed malignant are infrequent and may be unavoidable given overlapping imaging features of benign and malignant nodules. Knowledge of benign pathologies that can mimic malignancy may help reduce the incidence of unnecessary surgeries.

Establishing a clinical education track as part of a radiology residency is essential in shaping future radiology educators. Many obstacles will be encountered while starting, maintaining, and improving these educational pathways. Hurdles may include recruiting suitable residents for the track, recruiting and supporting faculty advisors, sustaining long-term resident engagement, counteracting educational exclusivity, and providing adequate time and financial support. Although every program and institution may face individualized "mountains" to overcome, they are not insurmountable. The goal of this review is to address different conflicts we have encountered while maintaining the clinical education tract at our institution and to provide tips for overcoming them.

OBJECTIVE:To quantify the association between computed tomography abdomen and pelvis with contrast (CTAP) findings and chest radiograph (CXR) severity score, and the incremental effect of incorporating CTAP findings into predictive models of COVID-19 mortality. METHODS:This retrospective study was performed at a large quaternary care medical center. All adult patients who presented to our institution between March and June 2020 with the diagnosis of COVID-19 and had a CXR up to 48 hours before a CTAP were included. Primary outcomes were the severity of lung disease before CTAP and mortality within 14 and 30 days. Logistic regression models were constructed to quantify the association between CXR score and CTAP findings. Penalized logistic regression models and random forests were constructed to identify key predictors (demographics, CTAP findings, and CXR score) of mortality. The discriminatory performance of these models, with and without CTAP findings, was summarized using area under the characteristic (AUC) curves. RESULTS:One hundred ninety-five patients (median age, 63 years; 119 men) were included. The odds of having CTAP findings was 3.89 times greater when a CXR score was classified as severe compared with mild (P = 0.002). When CTAP findings were included in the feature set, the AUCs for 14-day mortality were 0.67 (penalized logistic regression) and 0.71 (random forests). Similar values for 30-day mortality were 0.76 and 0.75. When CTAP findings were omitted, all AUC values were attenuated. CONCLUSIONS:The CTAP findings were associated with more severe CXR score and may serve as predictors of COVID-19 mortality.

To tune and test the generalizability of a deep learning-based model for assessment of COVID-19 lung disease severity on chest radiographs (CXRs) from different patient populations. A published convolutional Siamese neural network-based model previously trained on hospitalized patients with COVID-19 was tuned using 250 outpatient CXRs. This model produces a quantitative measure of COVID-19 lung disease severity (pulmonary x-ray severity (PXS) score). The model was evaluated on CXRs from 4 test sets, including 3 from the United States (patients hospitalized at an academic medical center (N = 154), patients hospitalized at a community hospital (N = 113), and outpatients (N = 108)) and 1 from Brazil (patients at an academic medical center emergency department (N = 303)). Radiologists from both countries independently assigned reference standard CXR severity scores, which were correlated with the PXS scores as a measure of model performance (Pearson R). The Uniform Manifold Approximation and Projection (UMAP) technique was used to visualize the neural network results. Tuning the deep learning model with outpatient data showed high model performance in 2 United States hospitalized patient datasets (R = 0.88 and R = 0.90, compared to baseline R = 0.86). Model performance was similar, though slightly lower, when tested on the United States outpatient and Brazil emergency department datasets (R = 0.86 and R = 0.85, respectively). UMAP showed that the model learned disease severity information that generalized across test sets. A deep learning model that extracts a COVID-19 severity score on CXRs showed generalizable performance across multiple populations from 2 continents, including outpatients and hospitalized patients.

PURPOSE/OBJECTIVE:To assess radiology representation, multimedia content, and multilingual content of United States lung cancer screening (LCS) program websites. MATERIALS AND METHODS/METHODS:We identified the websites of US LCS programs with the Google internet search engine using the search terms lung cancer screening, low-dose CT screening, and lung screening. We used a standardized checklist to assess and collect specific content, including information regarding LCS staff composition and references to radiologists and radiology. We also tabulated types and frequencies of included multimedia and multilingual content and patient narratives. RESULTS:We analyzed 257 unique websites. Of these, only 48% (124 of 257) referred to radiologists or radiology in text, images, or videos. Radiologists were featured in images or videos on only 14% (36 of 257) of websites. Radiologists were most frequently acknowledged for their roles in reading or interpreting imaging studies (35% [90 of 574]). Regarding multimedia content, only 36% (92 of 257) of websites had 1 image, 27% (70 of 257) included 2 or more images, and 26% (68 of 257) of websites included one or more videos. Only 3% (7 of 257) of websites included information in a language other than English. Patient narratives were found on only 15% (39 of 257) of websites. CONCLUSIONS:The field of Radiology is mentioned in text, images, or videos by less than half of LCS program websites. Most websites make only minimal use of multimedia content such as images, videos, and patient narratives. Few websites provide LCS information in languages other than English, potentially limiting accessibility to diverse populations.

The axilla contains several important structures which exist in a relatively confined anatomic space between the neck, chest wall, and upper extremity. While neoplastic lymphadenopathy may be among the most common axillary conditions, many other processes may be encountered. For example, expanded use of axillary vessels for access routes for endovascular procedures will increase the need for radiologists to access vessel anatomy, patency, and complications that may arise. Knowledge of axillary anatomy and pathology will allow the imager to systematically evaluate the axillae using various imaging modalities.