Radiomics-Enhanced Multi-task Neural Network for Non-invasive Glioma Subtyping and Segmentation

Zhiyuan Xue, Bowen Xin, Dingqian Wang, Xiuying Wang

Abstract: Non-invasive glioma subtyping can provide diagnostic support for pre-operative treatments. Traditional radiomics method for subtyping is based on hand-crafted features, so the capacity of capturing comprehensive features from MR images is still limited compared with deep learning method. In this work, we propose a radiomics enhanced multi-task neural network, which utilizes both deep features and radiomic features, to simultaneously perform glioma subtyping, and multi-region segmentation. Our network is composed of three branches, namely shared CNN encoder, segmentation decoder, and subtyping branch, constructed based on 3D U-Net. Enhanced with radiomic features, the network achieved 96.77% for two-class grading and 93.55% for three-class subtyping over the validation set of 31 cases, showing the potential in non-invasive glioma diagnosis, and achieved better segmentation performance than single-task network.