Self-supervised pretraining and network ensembling for spatial mapping of treatment-effect in recurrent GBM with physiologic MRI

Prior characterization of treatment-effect and tumor recurrence using deep learning approaches have not optimized for spatial classification within a single lesion, which could improve surgical planning and treatment. 10mm patches of pre-surgical anatomical and physiological images surrounding the locations of histopathologically-confirmed tissue samples were used to train our models. Including physiological images, pretraining on unlabeled data in an autoencoding task, and training with an alternative cross-validation approach that enabled many networks to be ensembled, we achieved an ensembled test AUROC of 0.814 and generated spatial maps of tumor probability and model uncertainty. Performance decreased when removing any of these components.