On validating multimodal MRI based stratification of IDH genotype in high grade gliomas using CNNs and its comparison to radiomics

Abstract

Radiomics based multi-variate models and state-of-art convolutional neural networks (CNNs) have demonstrated their usefulness for predicting IDH genotype in gliomas from multi-modal brain MRI images. However, it is not yet clear on how well these models can adapt to unseen datasets scanned on various MRI scanners with diverse scanning protocols. Further, gaining insight into the imaging features and regions that are responsible for the delineation of the genotype is crucial for clinical explainability. Existing multi-variate models on radiomics can provide the underlying signatures while the CNNs, despite better accuracies, more-or-less act as a black-box model. This work addresses these concerns by training radiomics based classifier as well as CNN classifier with class activation mapping (CAMs) on 147 subjects from TCIA and tests these classifiers directly and through transfer learning on locally acquired datasets. Results demonstrate higher adaptability of Radiomics with average accuracy of 75.4% than CNNs (68.8%), however CNNs with transfer learning demonstrate superior predictability with an average accuracy of 81%. Moreover, our CAMs display precise discriminative location on various modalities that is particularly important for clinical interpretability and can be used in targeted therapy.