Computerized Classification Method for Molecular Subtypes of Low-Grade Gliomas on Brain MR Images Using Modified ArcFace With Gram-Schmidt Orthogonalization

Abstract

Researchers have previously developed a computerized classification method for molecular subtypes of low-grade gliomas (LGGs) on brain magnetic resonance (MR) images using three-dimensional convolutional neural networks (3D-CNN) with softmax cross-entropy loss. However, a CNN with softmax cross-entropy loss cannot obtain sufficient discrimination power because it does not explicitly enforce any margin between each class label. Thus, in this study, we propose a new multiscale 3D-attention branch network (MS3D-ABN) using modified ArcFace, a deep metric learning method, with Gram-Schmidt orthogonalization for classifying molecular subtypes of LGGs. Our database included T1- and T2-weighted brain MR images of 217 patients from Tokyo Women's Medical University. LGGs included isocitrate dehydrogenase-wildtype astrocytomas (n=49), isocitrate dehydrogenase-mutant astrocytomas (n=58), and oligodendrogliomas (n=110). The MS3D-ABN using modified ArcFace with Gram-Schmidt orthogonalization learned how to make samples from feature vectors of the same class labels close to each other and further separate samples from different class labels to improve discrimination power. Gram-Schmidt orthogonalization played a role in maintaining orthogonality of feature vectors for each class label. The average classification accuracy of the proposed network was 66.8% (145/217), which was an improvement compared to MS3D-ABN without ArcFace (63.6%; 138/217) and MS3D-ABN with ArcFace (65.4%; 142/217).