brats-africa-eval
Topology-Driven Fusion of nnU-Net and MedNeXt for Accurate Brain Tumor Segmentation on Sub-Saharan Africa Dataset — Bohara et al. (2026) (arXiv:2604.15964, 2026)
What this evaluates
Evaluates the accuracy of deep learning models in segmenting brain tumor subregions and boundaries on low-field MRI scans from Sub-Saharan Africa. It probes the model's ability to handle regional imaging protocol limitations and topological deformations in medical image segmentation.
Datasets
- BraTS-Africa — total ?; splits: test (-1)
Metrics
Dice Similarity Coefficient (DSC) (primary) — range: [0, 1]
- DSC = 2|A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth segmentation masks. Ranges from 0 to 1, with 1 indicating perfect overlap.
Normalized Surface Distance (NSD) — range: [0, 1]
- NSD = 1 - (1/N) * Σ min(dist(a, B), tol) + min(dist(b, A), tol) / (2*tol), evaluated with a 1.0 mm tolerance. Ranges from 0 to 1, with 1 indicating perfect surface alignment.
Input / output format
Input: 3D low-field MRI scans of brain tumors from the Sub-Saharan Africa dataset.
Output: Pixel-wise segmentation masks for three tumor subregions: SNFH, NETC, and ET, plus a Legacy region mask.
Scoring recipe
def compute_dsc(pred, gt):
intersection = np.sum(pred & gt)
return 2.0 * intersection / (np.sum(pred) + np.sum(gt))
def compute_nsd(pred, gt, tol=1.0):
# Compute surface distances between pred and gt masks
# NSD = 1 - (1/N) * sum(min(d_a, tol) + min(d_b, tol)) / (2*tol)
# Returns value in [0, 1]
pass
Common pitfalls
- NSD is computed with a specific 1.0 mm tolerance, which must be explicitly set to match reported values.
- DSC is reported per subregion (SNFH, NETC, ET) and per anatomical grouping (Legacy, Lesion); averaging across regions without weighting can misrepresent performance.
- Pre-training on BraTS-2025 Task 1 is part of the evaluation protocol, not just training; models must be fine-tuned on BraTS-Africa before testing.
Evidence (verbatim from paper)
Our Baseline Model, nnU-Net 3D full resolution performed well overall on the BraTS-Africa dataset achieving high Dice scores for each tumor subregion: 0.930 for SNFH, 0.906 for NETC, and 0.906 for ET. The boundary alignment was demonstrated on NSD with a tolerance of 1.0 mm, for SNFH (0.830), NETC (0.827), and ET (0.894).
Citation
@misc{bohara2026topology,
title={Topology-Driven Fusion of nnU-Net and MedNeXt for Accurate Brain Tumor Segmentation on Sub-Saharan Africa Dataset},
author={Bohara et al. (2026)},
year={2026},
note={arXiv:2604.15964}
}
1---2name: brats-africa-eval3description: Evaluates the accuracy of deep learning models in segmenting brain tumor subregions and boundaries on low-field MRI scans from Sub-Saharan Africa. It probes the model's ability to handle regional imaging protocol limitations and topological deformations in medical image segmentation. Use when the user wants to benchmark on BraTS-Africa, or asks about evaluating this task. Reports Dice Similarity Coefficient (DSC).4---56# brats-africa-eval78> Topology-Driven Fusion of nnU-Net and MedNeXt for Accurate Brain Tumor Segmentation on Sub-Saharan Africa Dataset — Bohara et al. (2026) (arXiv:2604.15964, 2026)910## What this evaluates1112Evaluates the accuracy of deep learning models in segmenting brain tumor subregions and boundaries on low-field MRI scans from Sub-Saharan Africa. It probes the model's ability to handle regional imaging protocol limitations and topological deformations in medical image segmentation.1314## Datasets1516- **BraTS-Africa** — total ?; splits: test (-1)1718## Metrics1920- `Dice Similarity Coefficient (DSC)` **(primary)** — range: [0, 1]21 - DSC = 2|A ∩ B| / (|A| + |B|), where A and B are the predicted and ground truth segmentation masks. Ranges from 0 to 1, with 1 indicating perfect overlap.22- `Normalized Surface Distance (NSD)` — range: [0, 1]23 - NSD = 1 - (1/N) * Σ min(dist(a, B), tol) + min(dist(b, A), tol) / (2*tol), evaluated with a 1.0 mm tolerance. Ranges from 0 to 1, with 1 indicating perfect surface alignment.2425## Input / output format2627**Input**: 3D low-field MRI scans of brain tumors from the Sub-Saharan Africa dataset.2829**Output**: Pixel-wise segmentation masks for three tumor subregions: SNFH, NETC, and ET, plus a Legacy region mask.3031## Scoring recipe3233```python34def compute_dsc(pred, gt):35 intersection = np.sum(pred & gt)36 return 2.0 * intersection / (np.sum(pred) + np.sum(gt))3738def compute_nsd(pred, gt, tol=1.0):39 # Compute surface distances between pred and gt masks40 # NSD = 1 - (1/N) * sum(min(d_a, tol) + min(d_b, tol)) / (2*tol)41 # Returns value in [0, 1]42 pass43```4445## Common pitfalls4647- NSD is computed with a specific 1.0 mm tolerance, which must be explicitly set to match reported values.48- DSC is reported per subregion (SNFH, NETC, ET) and per anatomical grouping (Legacy, Lesion); averaging across regions without weighting can misrepresent performance.49- Pre-training on BraTS-2025 Task 1 is part of the evaluation protocol, not just training; models must be fine-tuned on BraTS-Africa before testing.5051## Evidence (verbatim from paper)5253> Our Baseline Model, nnU-Net 3D full resolution performed well overall on the BraTS-Africa dataset achieving high Dice scores for each tumor subregion: 0.930 for SNFH, 0.906 for NETC, and 0.906 for ET. The boundary alignment was demonstrated on NSD with a tolerance of 1.0 mm, for SNFH (0.830), NETC (0.827), and ET (0.894).5455## Citation5657```bibtex58@misc{bohara2026topology,59 title={Topology-Driven Fusion of nnU-Net and MedNeXt for Accurate Brain Tumor Segmentation on Sub-Saharan Africa Dataset},60 author={Bohara et al. (2026)},61 year={2026},62 note={arXiv:2604.15964}63}64```6566- arXiv: 2604.15964
qhjqhj00