⚡ electrical engineering

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

This paper presents a topology-driven fusion of nnU-Net and MedNeXt, enhanced by a refinement module and pre-training on BraTS 2025 data, to achieve accurate brain tumor segmentation on the challenging, low-quality MRI dataset from Sub-Saharan Africa.

Original authors: Prabin Bohara, Pralhad Kumar Shrestha, Arpan Rai, Usha Poudel Lamgade, Confidence Raymond, Dong Zhang, Aondona Lorumbu, Craig Jones, Mahesh Shakya, Bishesh Khanal, Pratibha Kulung

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Prabin Bohara, Pralhad Kumar Shrestha, Arpan Rai, Usha Poudel Lamgade, Confidence Raymond, Dong Zhang, Aondona Lorumbu, Craig Jones, Mahesh Shakya, Bishesh Khanal, Pratibha Kulung

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a map of a very tricky, foggy island (a brain tumor) using a satellite photo (an MRI scan). In rich countries, the photos are crystal clear, and the islands have well-defined borders. But in many parts of Sub-Saharan Africa, the "satellite photos" are often blurry, grainy, and taken with older, weaker cameras. Drawing the map manually is slow, tiring, and depends entirely on how good the artist (the doctor) is that day.

This paper is about a team of researchers who built a super-smart digital artist to draw these maps automatically, even when the photos are blurry.

Here is the story of how they did it, broken down into simple parts:

1. The Problem: The "Blurry Photo" Challenge

In many African countries, hospitals don't have the newest, most powerful MRI machines. The images they get are like looking at a painting through a foggy window.

The Issue: Standard computer programs (AI) are trained on clear, high-quality photos. When you show them a blurry photo, they get confused. They might draw the tumor too big, too small, or break it into tiny, disconnected pieces (like a puzzle where the pieces don't fit together).
The Goal: Create an AI that can see through the fog and draw the tumor boundaries perfectly, helping doctors treat patients faster.

2. The Tools: Two Different Artists

The researchers didn't just use one tool; they hired two different "digital artists" and asked them to work together.

Artist A (nnU-Net): This is the "Swiss Army Knife" of medical AI. It's famous because it automatically adjusts its own settings to fit whatever picture it sees. It's very reliable but sometimes gets the edges a little fuzzy.
Artist B (MedNeXt): This is a newer, specialized artist designed to understand the deep, complex structures of the brain. It sees details that Artist A might miss.

The Strategy: Instead of picking one, they made them work as a team. They took the drawings from both artists and blended them together. It's like asking two expert chefs to taste a soup and combine their best ideas to make the perfect flavor.

3. The Secret Sauce: The "Topology Refinement" (The Shape Fixer)

Here is the most creative part of the paper. Even with two great artists, the AI sometimes makes a weird mistake: Topological Errors.

The Metaphor: Imagine the tumor is a donut. A perfect AI should draw a donut with a hole in the middle. But a confused AI might draw a solid circle (no hole) or two separate circles (a broken donut). In medical terms, this is a "topological error." The shape is wrong, even if the size is close.
The Solution: The team added a third step called Topology Refinement. Think of this as a sculptor who comes in after the painters are done.
- The painters (nnU-Net and MedNeXt) do the rough work.
- The sculptor (the Topology Refinement module) looks at the painting and says, "Wait, this tumor shouldn't be broken in half," or "This part shouldn't be a solid block."
- The sculptor fixes the shape, ensuring the tumor looks like a real, continuous object, not a broken puzzle.

4. The Training: Learning from the Best

To teach their AI, they used a clever trick:

Step 1: They first taught the AI on a massive dataset of high-quality brain scans from around the world (the "BraTS 2025" data). This was like teaching a student with a textbook full of perfect diagrams.
Step 2: Then, they took that smart student and gave them a specific homework assignment: the blurry, difficult scans from Africa. This is called "fine-tuning." It's like taking a student who knows all the theory and putting them in a real-world internship to learn how to handle the messy, real-life situations.

5. The Result: A Better Map

The team tested their system and found:

The "blended" team of artists did a great job at identifying where the tumor was (measured by a score called Dice).
The "Sculptor" (Topology Refinement) didn't necessarily make the tumor bigger or smaller, but it made the edges much sharper and more accurate.
Most importantly, it fixed the weird "broken donut" errors, making the tumor look like a real, healthy biological structure.

Why Does This Matter?

In the real world, this means doctors in resource-limited areas can get a computer to help them draw the tumor boundaries in seconds instead of hours.

Faster Treatment: Patients don't have to wait days for a diagnosis.
Better Surgery: Surgeons can see exactly where the tumor starts and stops, even in blurry images, leading to safer operations.
Fairness: It brings high-level medical technology to places that usually can't afford it, helping to save lives in Sub-Saharan Africa.

In short: The researchers built a team of AI artists, taught them on the best data available, and added a "shape-fixing" sculptor to ensure the final map of the brain tumor is not just accurate in size, but perfect in shape.

1. Problem Statement

The paper addresses the critical challenge of automatic brain tumor segmentation in Low and Middle-Income Countries (LMICs), specifically Sub-Saharan Africa (SSA).

Context: Gliomas are aggressive tumors with high mortality rates in SSA due to limited medical infrastructure, a shortage of trained professionals, and delayed diagnoses.
Data Challenges: The BraTS-Africa dataset presents unique difficulties compared to high-income country datasets:
- Low-quality imaging: Scans are acquired on low-field MRI scanners with reduced contrast and high artifacts.
- Lack of protocols: Absence of defined national imaging protocols leads to diverse and inconsistent data.
- Topological Errors: Standard segmentation models often fail to preserve the correct topology of tumor structures (e.g., creating fragmented or smoothed boundaries), leading to poor performance on lesion-wise metrics despite high volumetric overlap.
Goal: To develop a robust automated segmentation pipeline that handles low-quality SSA data while ensuring topological correctness (preserving the shape and connectivity of tumor sub-regions).

2. Methodology

The authors propose a Topology-Driven Fusion Pipeline consisting of three main stages: Pre-training, Ensemble Modeling, and Topology Refinement.

A. Data Strategy

Target Dataset: BraTS-Africa 2025 (60 training cases, 35 validation cases) containing pre-operative adult glioma scans (T1, T2, T1c, T2-FLAIR).
Pre-training Strategy: To mitigate the scarcity and low quality of SSA data, the models were pre-trained on the BraTS 2025 Task 1 dataset (pre-treatment adult glioma from general sources) and then fine-tuned on the BraTS-Africa dataset.
Sub-regions: The task involves segmenting three specific tumor sub-regions:
1. Surrounding Non-Enhancing FLAIR Hyperintensity (SNFH).
2. Non-Enhancing Tumor Core (NETC).
3. Enhancing Tumor (ET).

B. Model Architecture & Ensemble

The pipeline fuses two state-of-the-art architectures:

nnU-Net (Baseline): A self-configuring U-Net variant. The authors trained both 2D slice-based and 3D full-resolution versions (6 encoder-decoder levels, deep supervision).
MedNeXt: A model based on ConvNeXt and 3D U-Net, featuring residual stages, depthwise convolutions, and Group Normalization.
Ensemble: The softmax predictions from the nnU-Net and MedNeXt models are combined using soft voting averaging to create a robust initial prediction.

C. Topology Refinement Module

To address topological errors (e.g., holes, disconnected components, or boundary deformations), the authors integrated a Universal Topology Refinement module as a post-processing step:

Mechanism: This module uses a trainable 3D U-Net (240×240×155) to correct the ensemble's output.
Training: It is trained on synthetic data where topology-perturbation masks (generated via orthogonal polynomial bases) are applied to introduce structural errors. This teaches the network to recognize and fix topological inconsistencies.
Objective: To improve boundary precision and capture thin structures that standard pixel-wise loss functions (like Dice) often smooth over.

3. Key Contributions

Domain Adaptation for LMICs: Demonstrated a successful transfer learning strategy (Pre-training on BraTS 2025 $\rightarrow$ Fine-tuning on BraTS-Africa) to handle low-field MRI data.
Topology-Aware Fusion: Proposed a novel pipeline that fuses nnU-Net and MedNeXt and explicitly corrects topological errors using a dedicated refinement network, moving beyond simple pixel-wise optimization.
Comprehensive Evaluation: Evaluated models not just on standard Dice Similarity Coefficient (DSC), but critically on Normalized Surface Distance (NSD) and Lesion-wise metrics, which are more sensitive to boundary accuracy and topological correctness.
Open Source: The code for the proposed models is made available to the community.

4. Results

The performance was evaluated on the BraTS-Africa validation set using Legacy (volume-based) and Lesion-wise metrics.

Dice Similarity Coefficient (DSC):
- The Fine-tuned nnU-Net achieved the highest Legacy DSC scores (SNFH: 0.936, NETC: 0.907, ET: 0.907).
- The Topology-aware model showed comparable DSC scores (SNFH: 0.920, NETC: 0.901, ET: 0.900) but did not significantly outperform the baseline in volumetric overlap.
Normalized Surface Distance (NSD@1.0mm):
- The Topology-aware model achieved the best NSD scores for NETC (0.829) and ET (0.895), indicating superior boundary alignment.
- However, for SNFH, the topology refinement resulted in a slight decline (0.810 vs 0.840 for fine-tuned), suggesting a trade-off between topological consistency and surface agreement for edematous regions.
Lesion-wise Performance: The topology refinement improved the preservation of small lesions and sharp edges, which is crucial for surgical planning, even if the overall volume overlap (Dice) remained similar.

5. Significance and Conclusion

Clinical Impact: The proposed method offers a viable solution for resource-constrained settings by automating tumor segmentation on low-quality MRI scans, potentially reducing the burden on radiologists and improving diagnostic throughput in SSA.
Methodological Insight: The study highlights that optimizing for topological correctness (via NSD and topology refinement) is distinct from optimizing for volumetric overlap (Dice). While the topology refinement module did not drastically increase Dice scores, it successfully improved boundary precision and structural integrity.
Future Directions: The authors acknowledge that the trade-off between surface refinement and volumetric accuracy needs further optimization. Future work should focus on developing topology-aware loss functions to better balance these metrics and achieve state-of-the-art performance specifically for brain tumor segmentation in diverse, low-resource environments.

In summary, this paper presents a robust, topology-driven ensemble approach that effectively adapts advanced deep learning models to the challenging conditions of Sub-Saharan African medical imaging, prioritizing structural accuracy alongside volumetric precision.