Oct 01

QP 28 - Radiation and Cancer Physics 14: Image-guided Planning, Novel Delivery Techniques and QA

1164 - Novel Cardiac Radiomics Using PET Imaging and Deep Learning to Predict Post-Radiotherapy Cardiac Toxicity for Lung Cancer Patients

12:15pm - 12:20pm PT

Room 160

Presenter(s)

Wenchao Cao, PhD - Thomas Jefferson University, Philadelphia, PA

W. Cao¹, M. Dichmann¹, N. Halder¹, E. Storozynsky², Y. Jia¹, N. L. Simone¹, M. Werner-Wasik³, A. P. Dicker¹, Y. Vinogradskiy¹, and W. Choi⁴; ¹Department of Radiation Oncology, Sidney Kimmel Medical College at Thomas Jefferson University, Philadelphia, PA, ²Department of Cardiology, Sidney Kimmel Medical College at Thomas Jefferson University, Philadelphia, PA, ³Department of Radiation Oncology, Sidney Kimmel Cancer Center, Philadelphia, PA, ⁴Dept of Radiation Oncology, Thomas Jefferson University Sidney Kimmel Cancer Center, Philadelphia, PA

Purpose/Objective(s):

Radiation-induced cardiac complications pose a significant concern in lung cancer radiotherapy. Conventional cardiotoxicity risk prediction models, which primarily rely on dose-volume parameters, remain primitive and lack accuracy. Fluorodeoxyglucose (FDG) Positron Emission Tomography (PET) scans, acquired as the standard of care for disease staging, can be repurposed to provide cardiac information. The objective of this study was to use deep learning (DL) based patient data synthesis methods to develop a novel radiomics signature using standard-of-care FDG PET scans to predict post-radiotherapy cardiotoxicity.

Materials/Methods:

We conducted a retrospective analysis of 100 consecutive lung cancer patients who underwent radiotherapy and had pre-treatment FDG PET-CT scans. Post-radiotherapy cardiotoxicity, including arrhythmia, ischemic events, heart failure, and cardiomyopathy, was assessed through chart review. The heart was contoured, and radiomic features were extracted from the cardiac PET signal. The predictive model evaluated the ability of three feature categories: (1) FDG-PET radiomic features, (2) clinical data (age, diabetes, hyperlipidemia, smoking status, hypertension, ECOG status), and (3) treatment-related factors (systemic therapy, maximum/mean heart dose, staging) to predict cardiotoxicity. The data were split 80%/20% training/test and assessed using accuracy and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). To address class imbalance, we developed a ß-variational autoencoder (VAE) and multi-layer perceptron (MLP) architecture. Model evaluation involved six machine learning classifiers (logistic regression, Support Vector Machines, random forests, XGBoost, LightGBM, and MLP), with Shapley Additive Explanations (SHAP) analysis employed to quantify feature importance and enhance model interpretability.

Results:

36 out of 100 (36%) patients developed cardiotoxicity. On the test set, standard modeling combining radiomics and clinical parameters predicted cardiotoxicity with an accuracy of 60.0%, precision of 63.6%, recall of 63.6%, F1-Score of 63.6%, and AUC-ROC of 61.6%. The ß-VAE-augmented SVM model demonstrated superior performance on the test set, achieving an accuracy of 70.0%, precision of 77.8%, recall of 63.6%, F1-score of 70.0%, and AUC-ROC of 61.6%. SHAP analysis revealed that the top 15% most influential predictors for cardiotoxicity included FDG-PET radiomic features (morphological characteristics, first-order energy, intensity patterns, and texture features), hypertension status, clinical stage, and mean and maximum heart doses.

Conclusion:

This study innovatively demonstrates that cardiac radiomics from repurposed standard-of-care PET scans, can serve as an early predictor of cardiotoxicity for lung cancer patients. Early prediction of cardiotoxicity can enable clinical decision support and allow for early cardiotoxicity mitigation strategies.