Improvement of Computerized Classification Method for Molecular Subtypes in Glioma on Brain MRI Images Using Multi-Scale 3D-Dual Attention Branch Networks with ElasticFace Loss

In this study, we developed a computerized classification method for molecular subtypes in LGGs (low-grade glioma) using multi-scale 3D-dual attention branch networks (MS3D-DABNs) with ElasticFace loss. Our database included T1-weighted, T2-weighted, and FLAIR brain MRI images from 217 patients (58 IDH-mutant astrocytomas, 49 IDH-wildtype astrocytomas, and 110 oligodendrogliomas) from Tokyo Women's Medical University. The proposed network consisted of a feature extractor, a dual attention branch combining spatial and channel attention modules, and a perception branch. The feature extractor first extracted the feature maps at different resolutions from each MRI image, which were then sequentially processed by the spatial and channel attention modules. The spatial attention module weighted features on the tumor region in the feature maps, while the channel attention module weighted important features in the feature maps. Finally, the perception branch classified LGGs into three molecular subtypes. During the training of the proposed network, ElasticFace loss was used to optimize intra-class compactness and inter-class separability. The average classification accuracy of the proposed method was 72.8% (Astrocytoma IDH-mutant: 65.5%, Astrocytoma IDH-wildtype: 71.4%, and Oligodendroglioma: 77.2%), showing an improvement compared to our previous network, multi-scale 3D spatial attention branch networks with ArcFace loss, 66.4% (51.7%, 61.2%, and 76.4%).