3577 - Association Between Cardiac Substructure Dose and Survival in NSCLC Patients Treated with Concurrent Chemoradiotherapy
Presenter(s)
G. Lin1,2, D. Niraula2, D. Dudas3, T. J. Dilling4, and I. El Naqa5; 1University of South Florida, Tampa, FL, 2Department of Machine Learning, Moffitt Cancer Center, Tampa, FL, 3Department of Radiation Oncology, LMU University Hospital, Munich, Germany, 4Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, 5Machine Learning & Radiation Oncology, Moffitt Cancer Center, Tampa, FL
Purpose/Objective(s): To determine which cardiac substructure radiation dose parameters independently predict overall survival (OS) in patients with non-small cell lung cancer (NSCLC) based on 3D treatment planning.
Materials/Methods:
We trained a multi-head deep survival network using 3D CT, 3D dose, and clinical variables (sex, age, delivery technique) from 360 NSCLC patients (IB–IIIC) treated with 60–70 Gy Survival times were discretized into five intervals using histogram-equalized binning. CT and dose volumes, masked by the whole heart, were processed through two parallel 3D convolutional networks, while clinical variables were encoded via a variational autoencoder. Latent features were concatenated and passed to a discrete survival head for OS prediction. Model interpretability was assessed using Grad-CAM. Cardiac substructure importance was quantified by overlaying activation maps onto segmentations generated from an open-source tool and an in-house cardiac atlas for 17 cardiac substructures then extracting the median activation per substructure for cohort-level comparison.Results: A 5-fold cross-validation on 70% of the dataset, repeated three times, produced an OS validation c-index of 0.786 ± 0.023 (stratified by intervals) and 0.792 ± 0.021 (non-stratified). On the 30% test set, c-indices were 0.771 ± 0.018 and 0.762 ± 0.019, respectively. Median Grad-CAM activation values from both the 3D dose and the CT showed long-tailed, right-skewed distribution across all cardiac substructures, with the CT branch exhibiting uniformly small values. Dose activation values overlapped substantially across substructures, indicating no single dominant substructure predictive of OS. Visual inspection revealed that activation hotspots aligned with high-dose regions along the heart wall. To explore this pattern, substructure activation values were grouped by primary tumor site and delivery technique. Upper-lobe tumors and 3D conformal plans consistently demonstrated higher median activation values. Formal hypothesis testing will be provided in the final presentation.
Conclusion: Preliminary findings show that no individual cardiac substructure serves as a unitary predictor of overall survival. Instead, consistently higher activation values in patients with upper-lobe tumors and those treated with 3D conformal radiotherapy suggest that radiation exposure to the whole heart, rather than any specific region, may drive survival risk. A comprehensive SHAP analysis incorporating both heart and lung dose will be provided in the final analysis.
| Dose | CT | |
| Valve Tricuspid | 0.0286 | 0 |
| CN Atrioventricular | 0.0281 | 0 |
| Ventricle L | 0.0266 | 0 |
| Ventricle R | 0.0259 | 0 |
| Atrium R | 0.0252 | 0 |
| A Coronary R | 0.0248 | 0.00097 |
| Valve Mitral | 0.0242 | 0 |
| A LAD | 0.0228 | 0.000002 |
| Heart | 0.0219 | 0 |
| A Cflx | 0.0215 | 0 |
| Valve Aortic | 0.0090 | 0 |
| Atrium L | 0.0061 | 0 |
| Valve Pulmonic | 0.0031 | 0 |
| A Coronary L | 0.0018 | 0 |
| CN Sinoatrial | 0.00003 | 0 |
| V Venacava S | 0 | 0 |
| A Pulmonary | 0 | 0 |
| A Aorta | 0 | 0 |