GU-Net: Diffuse Glioma Segmentation in Brain MRIs Using a Modified U-Net under Data Constraints

Abstract

Diffuse gliomas are a prevalent type of brain tumor in adults. Currently, treating these tumors is a time-consuming process. Radiologists manually identify and segment diffuse gliomas in Magnetic Resonance Images (MRIs), which are then used as reference by surgeons during treatment. Prior research conducted on automating this process utilizes machine learning (ML) models such as CNNs and U-Nets. One key piece of prior work is BU-Net, which slightly alters the architecture of U-Net. To contribute to this field, we propose a novel, simplified version of BU-Net, dubbed GU-Net, optimized specifically for low-computation neuroimaging. The proposed model is trained on a subset of the BraTS 2021 dataset, consisting of a mere 1647 images stemming from 549 different brain MRIs. Under data constraints, we achieve a 71.58% dice similarity coefficient (DSC) and 64.29% Intersection Over Union (IOU) on the testing dataset. Compared with U-Net's 0.672 and 0.611 and BU-Net's 0.613 and 0.554 on the same dataset, GU-Net’s success under data constraints compared to the other two models is shown. Our work specifically advances diagnosis in underprivileged areas and hospitals with less funding, as GU-Net requires less data to be used and has higher efficiency compared to existing solutions.