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

TL;DR

This study validates and deploys an AI system for automatic coronary segmentation and cardiac toxicity risk prediction, demonstrating its accuracy, clinical relevance, and potential to improve patient management in thoracic radiotherapy.

Contribution

The paper introduces a validated deep learning-based system for automated cardiac substructure segmentation and risk prediction, enabling large-scale surveillance and clinical decision support.

Findings

High geometric accuracy of AI segmentation (Dice > 0.87)

AI-derived dose metrics are associated with clinical outcomes

Deployment reduced high-risk dose exposure over time

Abstract

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.