Brain Tumor Segmentation in Sub-Sahara Africa with Advanced Transformer and ConvNet Methods: Fine-Tuning, Data Mixing and Ensembling

TL;DR

This study evaluates advanced deep learning architectures for brain tumor segmentation in Sub-Saharan Africa, demonstrating promising results with ensemble methods and emphasizing the importance of region-specific data for deploying AI in resource-limited settings.

Contribution

It compares state-of-the-art CNN and Transformer models on SSA datasets, highlighting the impact of data mixing, fine-tuning, and ensembling on segmentation performance.

Findings

MedNeXt achieved a Dice score of 0.84 for whole tumor

Ensembling models improved segmentation accuracy

Training on mixed datasets did not always enhance performance

Abstract

Brain tumors are among the deadliest cancers worldwide, with particularly devastating impact in Sub-Saharan Africa (SSA) where limited access to medical imaging infrastructure and expertise often delays diagnosis and treatment planning. Accurate brain tumor segmentation is crucial for treatment planning, surgical guidance, and monitoring disease progression, yet manual segmentation is time-consuming and subject to inter-observer variability. Recent advances in deep learning, based on Convolutional Neural Networks (CNNs) and Transformers have demonstrated significant potential in automating this critical task. This study evaluates three state-of-the-art architectures, SwinUNETR-v2, nnUNet, and MedNeXt for automated brain tumor segmentation in multi-parametric Magnetic Resonance Imaging (MRI) scans. We trained our models on the BraTS-Africa 2024 and BraTS2021 datasets, and performed validation on the BraTS-Africa 2024 validation set. We observed that training on a mixed dataset (BraTS-Africa 2024 and BraTS2021) did not yield improved performance on the SSA validation set in all tumor regions compared to training solely on SSA data with well-validated methods. Ensembling predictions from different models also lead to notable performance increases. Our best-performing model, a finetuned MedNeXt, achieved an average lesion-wise Dice score of 0.84, with individual scores of 0.81 (enhancing tumor), 0.81 (tumor core), and 0.91 (whole tumor). While further improvements are expected with extended training and larger datasets, these results demonstrate the feasibility of deploying deep learning for reliable tumor segmentation in resource-limited settings. We further highlight the need to improve local data acquisition protocols to support the development of clinically relevant, region-specific AI tools.