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

The real story here is not the model architecture but the dataset: MRI scans from Sub-Saharan Africa tend to be lower resolution and noisier than the Western hospital data most segmentation models train on, which means accuracy numbers from standard benchmarks do not transfer. Topology refinement is being used specifically to recover structural coherence that degrades when image quality drops.

This connects most directly to the SegWithU paper from April 16, which introduced uncertainty quantification for medical image segmentation as a way to flag where predictions are likely wrong. The two papers are attacking adjacent problems: SegWithU asks 'how confident should we be in a segmentation output,' while this work asks 'how do we produce a structurally valid output in the first place when input quality is poor.' Together they sketch a more complete pipeline for deploying segmentation in clinical settings where neither the scanner nor the model can be assumed to be ideal. The MADE benchmark from April 16 also reinforces a broader pattern: the medical AI community is increasingly building evaluation infrastructure around underrepresented populations and edge-case distributions rather than clean academic datasets.

Watch whether the BraTS Africa 2025 final leaderboard shows this ensemble holding its margin over single-model baselines across all tumor subregions, particularly the enhancing tumor class, which is most sensitive to image quality degradation.