147 - An Integrated AI Framework for Automated Cardiac Substructure Segmentation and Cardiotoxicity Risk Mitigation in Lung Cancer
Presenter(s)
X. Chen1, X. Zhang2, T. Xu3, R. Mohan2, Y. Zhao2, D. J. Rhee2, R. Lin4, M. Chen3, A. Ajdari5, S. Koutroumpakis6, N. L. Palaskas7, A. Deswal6, J. Niedzielski8, S. S. Shete9, L. E. Court2, B. Sun1, J. Yang2, and Z. Liao3; 1Department of Radiation Oncology, Dan L. Duncan Comprehensive Cancer Center, Baylor College of Medicine, Houston, TX, 2Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, 3Department of Thoracic Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 4Department of Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, TX, 5Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, 6Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 7Department of Cardiology, The University of Texas MD Anderson Cancer Center, Houston, TX, 8The University of Texas MD Anderson Cancer Center, Houston, TX, 9The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences, Houston, TX
Purpose/Objective(s): Radiation-induced cardiotoxicity remains a major long-term complication that significantly impacts the survivorship of patients with lung cancer. While cardiac substructure sparing is clinically desirable, its routine implementation is hindered by the labor-intensive nature of manual segmentation and the lack of integrated, actionable predictive tools. We developed and validated an end-to-end translational pipeline—integrating automated substructure segmentation, biomarker-based risk modeling, and personalized plan re-optimization—to bridge the gap between dosimetric research and clinical practice.
Materials/Methods: A customized nnU-Net architecture was trained to segment 19 cardiac substructures on non-contrast CTs from lung cancer patients (n=80). Accuracy was assessed via geometric indices (Dice) and clinical evaluation by two radiation oncologists and two cardiologists (n=42). This model was deployed on a retrospective cohort (n=160) to develop a logistic regression model predicting post-treatment elevation of high-sensitivity cardiac troponin T (hs-cTnT), a validated early marker of myocardial injury. Predictors included clinical factors, substructure dose-volume histogram (DVH) metrics, and radiomic/dosiomic features. The model underwent independent validation in a prospective cohort (n=57). Finally, an automated "screen-and-steer" workflow was tested on 35 patients to identify high-risk cases and perform model-informed plan re-optimization by establishing substructure-based objective functions and dynamic weight-tuning.
Results: The auto-segmentation model achieved excellent performance, with an average Dice of 0.81 (including coronary arteries) and a 94% clinical acceptability rate. The substructure-specific DVH model significantly outperformed whole-heart and radiomic/dosiomic models in both cross-validation (AUROC: 0.71, 95% CI: 0.70–0.73) and prospective validation (AUROC: 0.60 vs. =0.51). The left anterior descending coronary artery (LAD) V20Gy and right ventricle (RV) maximum dose emerged as the most critical predictors of injury. In the re-optimization phase, the pipeline identified 7/35 high-risk patients; automated re-planning achieved a mean reduction in LAD V20Gy of 15.5% and RV maximum dose of 9 Gy without compromising PTV coverage or other OAR constraints.
Conclusion: This study demonstrates the feasibility and translational value of an end-to-end, automated framework for personalized cardiotoxicity risk management in lung cancer radiotherapy. By integrating cardiac substructure auto-segmentation, biomarker-driven risk prediction, and plan re-optimization, the proposed approach enables clinically actionable cardiac sparing at scale. Importantly, this tool is currently being tested for implementation in low-resource community hospital settings, supporting its potential for broad clinical adoption and impact.