Introduction
Accurate brain tumor (glioma) segmentation is a critical component of neuro-oncological treatment planning. However, manual segmentation is a labor-intensive process requiring specialized expertise, which is often in short supply in regions like Sub-Saharan Africa (SSA). While deep learning has shown promise in automating this task, most existing models are trained on high-quality datasets from the Global North. These models often fail to generalize to the lower-quality MRI scans—frequently produced by low-field magnets—and the unique patient demographics found in SSA clinical settings.
EMedNeXt addresses these challenges by adapting the MedNeXt V2 architecture specifically for the constraints of low-resource environments. By employing a larger field of view, boundary-aware loss functions, and a "structured freezing" fine-tuning strategy, the framework achieved state-of-the-art results on the BraTS-Lighthouse 2025 challenge.
Methodology
The core of the EMedNeXt framework is MedNeXt V2, which evolves the standard MedNeXt architecture by integrating ConvNeXt-style blocks into a 3D U-Net structure. Several key enhancements were introduced to improve robustness:
- Architectural Upgrades: The authors unified kernel sizes to a fixed $3 \times 3 \times 3$ and increased the input patch size (Field of View) to $160 \times 160 \times 128$. This larger patch size allows the network to capture more global context, which is essential for orienting the tumor within the brain when image quality is degraded.
- Deep Supervision: Auxiliary segmentation heads were integrated at multiple decoder levels to facilitate better gradient flow during training.
- Boundary-Aware Loss: To improve the precision of tumor edges, the researchers combined standard Dice-Focal loss with a 3D Sobel-based boundary term. This forces the model to prioritize the difficult-to-define boundaries of the lesion.
Domain-Adaptive Training Strategy
To overcome the extreme scarcity of annotated data from the SSA region (only 60 training cases), the authors implemented a two-phase training approach:
- Pre-training: The model was first trained on the PPTAG (Pre- and Post-Treatment Adult Glioma) dataset, which contains 1,195 high-quality MRI scans. This phase allowed the model to learn a broad distribution of general tumor features.
- Structured Fine-tuning: The model was then fine-tuned on the BraTS-SSA dataset. Crucially, the encoder was frozen during this phase. This "structured freezing" strategy prevents overfitting on the small SSA sample while allowing the decoder to adapt to the specific noise and artifacts characteristic of the African clinical data.
Results and Discussion
The EMedNeXt framework demonstrated significant performance gains over the MedNeXt V1 baseline. On the hidden validation set of the BraTS-Lighthouse 2025 challenge, the model achieved an average Lesion-Wise Dice Similarity Coefficient (DSC) of 0.897.
The model was particularly effective at segmenting the "Whole Tumor" (WT) class, reaching a Dice score of 0.933. Furthermore, the inclusion of the Sobel operator significantly improved the Normalized Surface Dice (NSD) scores, indicating high accuracy in tracing tumor boundaries. Post-processing steps, including connected-component pruning and hierarchical enforcement (ensuring the Enhancing Tumor is nested within the Tumor Core and Whole Tumor), further refined the final outputs.
Conclusion
EMedNeXt provides a robust solution for brain tumor segmentation in data-scarce and low-resource environments. By combining advanced architectural features with a domain-adaptive training strategy, the framework bridges the performance gap between high-quality research datasets and real-world clinical data from Sub-Saharan Africa. Future work will focus on extending this framework to pediatric populations and exploring newer backbones to further enhance segmentation accuracy.
Summary updated