Author: Wookjin Choi, Michael Dichmann, Adam Dicker, Nilanjan Haldar, Yingcui Jia, Nicole L Simone, Eugene Storozynsky, Yevgeniy Vinogradskiy, Maria Werner-Wasik 👨🔬
Affiliation: Thomas Jefferson University, 9Department of Radiation Oncology, Thomas Jefferson University 🌍
Purpose: Cardiotoxicity remains a significant limitation for lung cancer patients treated with radiotherapy. Pre-radiotherapy cardiac conditions increase the probability of patients developing cardiotoxicity after treatment. Automated methods are needed to identify patients with pre-radiotherapy cardiac conditions. FDG PET scans are acquired as standard of care for disease staging and can be repurposed to provide cardiac information. The purpose of this work was to develop a novel radiomics signature using standard-of-care FDG PET scans to predict pre-radiotherapy cardiac conditions.
Methods: This study included 100 lung cancer patients consecutively considered for radiotherapy with pre-treatment FDG PET-CT scans. Pre-radiotherapy cardiac conditions were evaluated using chart review and included arrhythmias, ischemic events, heart failure, and cardiomyopathy. The heart was contoured, and radiomics features were extracted from the PET signal in the heart. Machine learning models used cardiac PET-based radiomics to predict pre-radiotherapy cardiac conditions. The data were split 80%/20% training/validation. Wilxcon testing and Hierarchical clustering were applied for feature reduction, followed by Recursive Feature Elimination. Prediction models were trained and evaluated with 10x10 cross-validation and independently validated on the validation set. The ability of the radiomics model was assessed using accuracy and area under the ROC curve (AUC).
Results: 22% of the patients had pre-radiotherapy cardiac conditions. The best-performing model was an artificial neural network (ANN) with the top 10 features, including 1st-order statistics, shape, texture, and wavelet features. The ANN cardiac PET radiomics model predicted pre-existing cardiac conditions with 83.3±10.4% (mean ± standard deviation) accuracy and 0.93±0.08 AUC on the training data and 72.8±4.5% accuracy and 0.86±0.02 AUC on the validation data.
Conclusion: This study demonstrated the potential of cardiac PET signals to predict pre-radiotherapy cardiac conditions. Our work repurposes standard-of-care PET scans to evaluate the heart and can be used to improve the prediction of patients at high risk for post-radiotherapy cardiotoxicity.