Topology-Driven Fusion of nnU-Net and MedNeXt for Accurate Brain Tumor Segmentation on Sub-Saharan Africa Dataset

A multi-institutional research team has published a novel AI architecture that fuses two leading medical imaging models—nnU-Net and MedNeXt—with a custom topology refinement module. The work, submitted for the Brain Tumor Segmentation (BraTS) Africa 2025 Challenge, specifically tackles the acute problem of automating brain tumor analysis in Low and Middle-Income Countries (LMICs). These regions face unique hurdles, including the widespread use of low-field MRI scanners, a lack of standardized imaging protocols, and limited healthcare resources, which degrade data quality and challenge standard AI models.

The team's key innovation is a topology-driven fusion approach that corrects for deformation errors in model predictions. They pre-trained their system on the higher-quality BraTS 2025 glioma dataset before fine-tuning it on the challenging BraTS-Africa data. This strategy, combined with the new refinement module, yielded impressive results: a Normalized Surface Distance (NSD) score of 0.895 for segmenting the Enhancing Tumor (ET) region, and scores of 0.810 and 0.829 for other tumor sub-regions. These metrics represent a significant accuracy boost for low-quality scans.

This research demonstrates a practical pathway for adapting state-of-the-art AI to real-world, resource-constrained clinical environments. By directly addressing the 'topological errors' that arise from poor image quality, the model provides more reliable and anatomically plausible segmentations. This work is a crucial step toward deploying equitable AI diagnostic tools that can function effectively outside well-funded, research-oriented hospitals.