EMedNeXt: An Enhanced Brain Tumor Segmentation Framework for Sub-Saharan Africa using MedNeXt V2 with Deep Supervision

TL;DR

This work targets robust brain tumor segmentation in sub-Saharan Africa where MRI data and radiology resources are limited. It introduces EMedNeXt, a MedNeXt V2–based pipeline with deep supervision, a larger region of interest, SSA-specific fine-tuning, model ensembling, and a boundary-aware post-processing scheme. By pretraining on the PPTAG dataset and fine-tuning on the SSA BraTS data, the approach achieves strong performance (LesionWise DSC $=0.897$, NSD $=0.541$ at $0.5\mathrm{~mm}$ and $0.84$ at $1.0\mathrm{~mm}$) on a hidden SSA validation set, demonstrating robustness to domain shifts and resource constraints. The combination of data fusion, transfer learning, ensembling, and topology-aware post-processing has practical implications for equitable, automated tumor quantification in under-resourced healthcare settings and can be extended to pediatric data and new architectures.

Abstract

Brain cancer affects millions worldwide, and in nearly every clinical setting, doctors rely on magnetic resonance imaging (MRI) to diagnose and monitor gliomas. However, the current standard for tumor quantification through manual segmentation of multi-parametric MRI is time-consuming, requires expert radiologists, and is often infeasible in under-resourced healthcare systems. This problem is especially pronounced in low-income regions, where MRI scanners are of lower quality and radiology expertise is scarce, leading to incorrect segmentation and quantification. In addition, the number of acquired MRI scans in Africa is typically small. To address these challenges, the BraTS-Lighthouse 2025 Challenge focuses on robust tumor segmentation in sub-Saharan Africa (SSA), where resource constraints and image quality degradation introduce significant shifts. In this study, we present EMedNeXt -- an enhanced brain tumor segmentation framework based on MedNeXt V2 with deep supervision and optimized post-processing pipelines tailored for SSA. EMedNeXt introduces three key contributions: a larger region of interest, an improved nnU-Net v2-based architectural skeleton, and a robust model ensembling system. Evaluated on the hidden validation set, our solution achieved an average LesionWise DSC of 0.897 with an average LesionWise NSD of 0.541 and 0.84 at a tolerance of 0.5 mm and 1.0 mm, respectively.

Figures (6)

• Figure 1: Cross sections of the four modalities obtained from a sample data-point from the SSA dataset along with the corresponding segmentation masks
• Figure 2: Cross sections of the four modalities obtained from a sample data-point from the PPTAG dataset along with the corresponding segmentation masks.
• Figure 3: MedNeXt V2: High-level architecture. A 4-channel 3D MRI input ($160\times160\times128$) is processed through a stem block, followed by an encoder that extracts hierarchical features, a bottleneck, and a decoder that reconstructs segmentation maps.
• Figure 4: Detailed MedNeXt V2 architecture: (a) encoder path, (b) decoder path, and (c) low-level block components.
• Figure 5: EMedNeXt: Training Flow