Brain Tumour Segmentation for Sub-Saharan African Population Using 3D UNet

Abdulquddus Adedayo Ajibade, Kenneth Aguh, Gabriel Babatunde, Adaobi Chiazor Emegoakor, Maruf Adewole

Keywords: Attention module, Wavelets, UNet, Segmentation, BraTS-Africa, Glioma, MRI

TL;DR: Deep learning UNet model with attention & wavelets improves brain tumor segmentation on unique SSA dataset (Dice Score: 0.8355).

Abstract: Patients diagnosed with Glioma generally have a survival rate of less than two years. The severity of the disease is even worse in low-resource settings such as Sub-Saharan Africa (SSA). Glioma in SSA is characterized with advanced stage presentation and large tumour volume. The Brain tumour segmentation (BraTS) Challenge featured dataset from SSA for the first time ever since running for more than a decade. The peculiarity of the location where these data were obtained finds its way into the dataset. Issues such as image quality, noise, unusually large tumours and low number of samples in the dataset, this makes the dataset different from what is obtainable in developed regions. Dataset imbalance also exists among the labelling of the sub-regions of the brain tumour with Peritumoral Oedema (OD) having larger labelling compared to Enhancing Tumour (ED) and lastly Necrotic Core (NC). This study aims to address the labelling imbalance, low number of samples and the unusual characteristic features of BraTS-Africa da taset. Here, we implemented UNet to develop African brain tumour segmentation model. An average Dice Score Coefficient (DSC) of 0.8355. This study suggests the effectiveness of attention module and wavelets in improving the performance of UNet. Despite the peculiarity of Dataset obtained from SSA, A deep learning model can be developed to address the diagnostic need of the region.

Submission Category: Machine learning algorithms

Submission Number: 71