Clinical Validation and Prospective Deployment of an Automated Deep Learning-Based Coronary Segmentation and Cardiac Toxicity Risk Prediction System

Importance: Coronary algorithm for cardiac sub structures and prospective real-time surveillance of cardiac dose exposure. Methods: Retro and prospective study to validate AI auto-segmentation. A 3D UNet was trained on 560 thoracic CT scans from a single institution (2003-2014) and validated internally (n=70). External validation was performed in 283 patients treated at an independent institution (2005-2020). Clinical implementation comprised (1) retrospective analysis of 3,399 lung cancer patients treated in 2014-2022 and (2) prospective surveillance of 1,386 consecutive patients in 2023. Geometric accuracy, concordance of dose-volume parameters; association of AI-derived substructure metrics with outcome; temporal dose trends; and the proportion of patients exceeding prespecified risk. Results: Median (inter-quartile range) Dice/ASSD were 0.95 (0.94-0.96)/1.1 mm for the heart and 0.87 (0.82-0.90)/1.9 mm for the LAD; the median absolute difference between AI and manual LAD V15 was 1%. AI-derived LAD V15 remained independently associated with MACE (sub distribution hazard ratio [HR], 1.03%; 95% CI, 1.01-1.05) and ACM (adjusted HR, 1.02; 95% CI, 1.00-1.03), internally and externally. Retrospective deployment showed a 32% relative decline in median LAD V15 from 2014 to 2022 (12% to 8%) and identified high- risk doses in 1,086 of 3,399 patients (32%). Prospective surveillance flagged 264 contemporary patients (19%) for potential cardiology referral. Conclusions: A validated AI system accurately segments cardiac substructures, reproduces dose-outcome relationships, enables large-scale surveillance, and point-of-care alerts for high-risk patients. Automated cardiac dose monitoring could facilitate adoption of coronary-sparing therapy and follow-up.