Computerized classification method for molecular subtypes of low-grade glioma in brain MRI images using 3D-attention branch networks with additive angular margin penalty

Abstract

The purpose of this study was to develop a computerized classification method for molecular subtypes of low-grade gliomas (LGGs) in brain MRI (magnetic resonance imaging) images using multi-scale three-dimensional attention branch networks (MS3D-ABNs) with an additive angular margin penalty. Our database consisted of brain T1-weighted, T2-weighted, and FLAIR MRI images for 217 patients (58 IDH-mutant astrocytomas, 49 IDH-wildtype astrocytomas, and 110 oligodendrogliomas). The proposed network was constructed from a feature extractor, an attention branch, and a perception branch with an additive angular margin penalty. The feature extractor first extracted the feature maps of different resolutions from brain T1-weighted, T2-weighted, and FLAIR MRI images, respectively. The attention branch generated attention maps focusing on a tumor region. The feature maps were then multiplied by the attention maps to weight features on the tumor region in the feature maps. The perception branch finally evaluated the molecular subtypes of LGGs by determining the cosine similarity between the feature vector obtained from applying a global average pooling to the feature map and the representative vector of each molecular subtype class. In training the proposed network, an angular margin penalty was added to the angle between the feature vector of input image and the representative vector of the same class as the input image to make those vectors to be closer each other. The classification accuracy for the proposed network was 66.4%, showing an improvement when compared to the MS3D-ABNs without the additive angular margin penalty (60.4%).