Episode 19: Leveraging Federated Learning to Supplement an AI Learning Dataset

Episode 24: When do we actually need to measure lung shunt fraction before yttrium-90 liver therapy?

This episode covers a study from the Mallinckrodt Institute of Radiology at Washington University in St. Louis evaluating whether patient-specific lung shunt fraction measurement is necessary for every yttrium-90 selective internal radiation therapy case. Across 354 cases, the authors propose a new pretreatment metric called LSFbound — derived from liver mass, lung mass, and dose thresholds — that identifies which patients can safely skip the macroaggregated albumin imaging workflow. For most cases without large tumors or macrovascular invasion, the lung shunt simply does not constrain treatment planning, offering a path to a streamlined, single-session workflow. When is patient-specific lung shunt fraction necessary in 90Y selective internal radiation therapy of liver cancer? Thomas et al. Radiology Advances, 2026, 3, umag007. [https://doi.org/10.1093/radadv/umag007]

Episode 23: Predicting severe pancreatitis from admission CT with deep learning

This episode discusses a study from New York University evaluating whether deep learning can predict acute pancreatitis severity from contrast-enhanced CT acquired within 24 hours of admission. Using self-supervised pretraining on about 12,000 unlabeled scans followed by supervised fine-tuning, the model achieved an AUROC near 0.89 for severe pancreatitis on both an internal NYU test set and an external multicenter Hungarian cohort of 518 patients, outperforming traditional clinical and imaging-based scoring systems. The work suggests that opportunistic AI triage on routinely acquired CT could support earlier, more accurate risk stratification in the emergency department. Deep learning-based prediction of acute pancreatitis severity from abdominal CT with multicenter external validation. Xu et al. Radiology Advances, 2026, 3, umag020 [https://doi.org/10.1093/radadv/umag020]

Episode 22: Can LLM-generated summaries help patients understand lung cancer screening reports?

This episode discusses a study from the University of California, San Francisco in the United States that tested whether GPT-4o-generated patient-friendly summaries improve comprehension of lung cancer screening CT reports. In a within-subjects survey of 1,815 adults across Lung-RADS 1, 2S, and 4B vignettes, the summaries significantly improved objective comprehension and reduced anxiety for all three report types. Largest gains were in participants with low self-rated English and health literacy. These findings support using LLM summariesas a potential health-equity tool, while highlighting the unmet patient need for personalized next-steps guidance. Self-reported comprehension of large language model-generated summaries of lung cancer screening reports: a vignette survey. Serna et al. Radiology Advances, 2026, 3, umag008. [https://doi.org/10.1093/radadv/umag008]

Episode 21: Can AI catch cardiomegaly on chest CTs ordered for other reasons?

This episode explores a study from the University of Texas Southwestern Medical Center and MD Anderson Cancer Center in the United States that clinically validates an FDA-cleared AI tool for measuring total cardiac volume on non-contrast, non-gated chest CT. Across 307 patients with paired echocardiography, the AI discriminated normal from abnormal cardiac volume with an AUC of 0.81 in men and 0.77 in women, and far outperformed routine radiologist sensitivity for cardiomegaly. The tool offers a tunable, reproducible opportunistic screening layer on chest CT's already being performed. Radiology Advances, 2026, 3, umag013. Fan et al. [https://doi.org/10.1093/radadv/umag013]

Episode 20: Minimum Data for Maximum Accuracy

This episode explores a study from the Emory Sports Performance and Research Center and the University of Lausanne that determined how few annotated MRI exams are needed to train a reliable deep learning model for thigh muscle segmentation. Using the nnU-Net framework with incrementally larger training sets, the researchers found that just 20 high-quality annotated subjects produced clinically acceptable segmentation across 14 thigh muscles, with biomarker agreement virtually indistinguishable from expert manual segmentation. All tools and trained models have been made openly available. Optimizing MRI annotation workflows for high-accuracy deep learning thigh muscle segmentation in athletes. Slutsky-Ganesh et al. Radiology Advances, 2026, 3, umag005 [https://doi.org/10.1093/radadv/umag005]