Bridging the Gap: Generalising State-of-the-Art U-Net Models to Sub-Saharan African Populations

• DOI: 10.48550/arxiv.2312.11770
• Publication Date: 2023-01-01
• Authors: Alyssa R. Amod; Alexandra Smith; Pearly Joubert; Confidence Raymond; Dong Zhang; Udunna Anazodo; YD Motchon; Tinashe Mutsvangwa; Sébastien Quetin
• Keywords: Training set

A critical challenge for tumour segmentation models is the ability to adapt diverse clinical settings, particularly when applied poor-quality neuroimaging data. The uncertainty surrounding this adaptation stems from lack of representative datasets, leaving top-performing without exposure common artifacts found in MRI data throughout Sub-Saharan Africa (SSA). We replicated a framework that secured 2nd position 2022 BraTS competition investigate impact dataset composition on model performance and pursued four distinct approaches through training with: 1) BraTS-Africa only (train_SSA, N=60), 2) BraTS-Adult Glioma (train_GLI, N=1251), 3) both datasets together (train_ALL, N=1311), 4) further train_GLI with (train_ftSSA). Notably, smaller low-quality alone (train_SSA) yielded subpar results, larger high-quality (train_GLI) struggled delineate oedematous tissue validation set. most promising approach (train_ftSSA) involved pre-training neuroimages then fine-tuning it smaller, dataset. This outperformed others, ranking second MICCAI global external testing phase. These findings underscore significance sample sizes broad improving performance. Furthermore, we demonstrated there potential such by them wider range locally.