Prediction of Survival During Immunotherapy of Recurrent High-Grade Glioma Using End-to-End CNN Deep Learning Versus Radiomics Models

Motivation: Markers to predict overall survival (OS) in recurrent high-grade glioma (HGG) patients undergoing immunotherapy, are needed to help select patients for trials and personalize treatment. Goal(s): To evaluate and compare deep learning and radiomic models for OS prediction in recurrent HGG, using manual segmentation, automated segmentation, and an end-to-end deep learning approach. Approach: We developed a segmentation-free CNN model, and radiomics models using features extracted from manually and CNN-segmented ROIs. We compared accuracy of OS prediction in 154 patients. Results: Radiomics from manual segmentation was more accurate than automated methods. The end-to-end CNN model achieved similar performance to the robust-feature manual-segmentation-based radiomics model. Impact: End-to-end CNN models can produce similar accuracy in recurrent HGG patient survival prediction during immunotherapy, compared to robust-feature radiomics from manual segmentation, and may add value in initial patient selection for immunotherapy trials, and personalization of therapy.