arxiv: 2604.15964 · v1 · submitted 2026-04-17 · 📡 eess.IV · cs.CV· cs.LG

Recognition: unknown

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

Prabin Bohara , Pralhad Kumar Shrestha , Arpan Rai , Usha Poudel Lamgade , Confidence Raymond , Dong Zhang , Aondona Lorumbu , Craig Jones

show 3 more authors

Mahesh Shakya Bishesh Khanal Pratibha Kulung

Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:45 UTC · model grok-4.3

classification 📡 eess.IV cs.CVcs.LG

keywords brain tumor segmentationtopology refinementnnU-NetMedNeXtBraTS-Africamedical image segmentationlow-resource MRIdeep learning fusion

0 comments

The pith

Adding a topology refinement module to fused nnU-Net and MedNeXt models raises Normalized Surface Distance scores for brain tumor segmentation on lower-quality African MRI data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses the difficulty of accurate automatic brain tumor segmentation in low-resource settings, where MRI scans vary in quality because of inconsistent protocols, low-field scanners, and limited infrastructure. The authors pre-train nnU-Net and MedNeXt on the higher-quality BraTS 2025 adult glioma dataset, fine-tune on the BraTS-Africa collection, and insert a topology refinement module to fix deformations caused by topological errors in the model outputs. This produces reported NSD improvements to 0.810 on surrounding non-enhancing FLAIR hyperintensity, 0.829 on non-enhancing tumor core, and 0.895 on enhancing tumor. A sympathetic reader would care because reliable segmentation directly supports diagnosis and treatment planning where healthcare resources are scarce.

Core claim

The paper establishes that fusing nnU-Net and MedNeXt after pre-training on BraTS 2025 data, then applying a topology refinement module, corrects topological errors and deformations in predictions on the BraTS-Africa dataset, delivering Normalized Surface Distance values of 0.810 for SNFH, 0.829 for NETC, and 0.895 for ET.

What carries the argument

The topology refinement module, which detects and corrects topological errors and resulting deformations in the segmentation predictions produced by the fused nnU-Net and MedNeXt models.

If this is right

The fused model with the refinement module produces higher boundary accuracy, as measured by NSD, across the three tumor subregions.
Pre-training on the BraTS 2025 dataset enables effective fine-tuning despite the domain shift to lower-quality African scans.
The reported NSD values indicate improved handling of topological inconsistencies that commonly arise in low-field MRI segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Topology refinement may serve as a lightweight post-processing step for other segmentation architectures facing similar domain-shift problems.
Direct integration of topological constraints into the training objective could reduce reliance on a separate refinement stage.
The method could extend to other low-resource medical imaging tasks where image quality varies across sites.

Load-bearing premise

The topology refinement module corrects topological errors without creating new inaccuracies, and pre-training on BraTS 2025 data transfers usefully to the lower-quality BraTS-Africa scans despite differences in scanners and protocols.

What would settle it

A direct comparison experiment in which adding the topology refinement module to the fused model produces no gain or a drop in NSD scores on the BraTS-Africa test set would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.15964 by Aondona Lorumbu, Arpan Rai, Bishesh Khanal, Confidence Raymond, Craig Jones, Dong Zhang, Mahesh Shakya, Prabin Bohara, Pralhad Kumar Shrestha, Pratibha Kulung, Usha Poudel Lamgade.

**Figure 1.** Figure 1: Topological error visualization of reference standard segmentation and predicted output before topology refinement, where yellow color denotes prediction error. lead to poor generalization when applied to different segmentation tasks. To address this challenge, Liu et al. proposed a universal topology preservation and refinement method [20]. This method creates topology-perturbation masks using randomly … view at source ↗

**Figure 2.** Figure 2: Model of our proposed Pipeline using an Ensemble of nnU-Net and MedNeXt with Topology-aware Post-Processing 2.3.3 Topology Refinement Model : We adopted the Universal Topology Refinement approach to address topological error correction [20], which generated synthetic segmentation labels and also provided a trainable pipeline network [19] to reduce topological errors in the baseline model. We used the topo… view at source ↗

**Figure 3.** Figure 3: Visualization of multimodal MRI inputs (T1c, T1w, T2-FLAIR, T2w) for cases BraTS-SSA-00010-000, BraTS-SSA-00025-000, BraTS-SSA-00096-000, along with corresponding reference standard segmentation masks, and baseline nnU-Net predicted tumor segmentation results. The segmentation overlays highlight tumor subregions including non-enhancing tumor (red), enhancing tumor (blue), and surrounding non-enhancing FL… view at source ↗

**Figure 4.** Figure 4: Multiplanar visualization (Axial, Coronal, and Sagittal views) of a brain MRI scan from the BraTS-Africa-00015 case. The first column shows the original T1c MRI slices. The second column overlays the ground truth tumor segmentation mask on the MRI, while the third, fourth and fifth columns overlay the predicted segmentation masks from the nnU-Net 3D, nnU-Net 2D, and MedNeXt models, respectively. This layo… view at source ↗

read the original abstract

Accurate automatic brain tumor segmentation in Low and Middle-Income (LMIC) countries is challenging due to the lack of defined national imaging protocols, diverse imaging data, extensive use of low-field Magnetic Resonance Imaging (MRI) scanners and limited health-care resources. As part of the Brain Tumor Segmentation (BraTS) Africa 2025 Challenge, we applied topology refinement to the state-of-the-art segmentation models like nnU-Net, MedNeXt, and a combination of both. Since the BraTS-Africa dataset has low MRI image quality, we incorporated the BraTS 2025 challenge data of pre-treatment adult glioma (Task 1) to pre-train the segmentation model and use it to fine-tune on the BraTS-Africa dataset. We added an extra topology refinement module to address the issue of deformation in prediction that arose due to topological error. With the introduction of this module, we achieved a better Normalized Surface Distance (NSD) of 0.810, 0.829, and 0.895 on Surrounding Non-Enhancing FLAIR Hyperintensity (SNFH) , Non-Enhancing Tumor Core (NETC) and Enhancing tumor (ET).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper fuses nnU-Net and MedNeXt, pre-trains on BraTS 2025 then fine-tunes on BraTS-Africa, and adds a topology module that claims better NSD, but supplies no baselines or ablations to show the module's actual effect.

read the letter

The main takeaway is that this is a straightforward application paper for the BraTS-Africa 2025 challenge. The authors combine two established segmentation networks, pre-train on the standard adult glioma data, fine-tune on the lower-quality Sub-Saharan Africa scans, and insert a topology refinement step to reduce deformed predictions. They report NSD values of 0.810, 0.829, and 0.895 on SNFH, NETC, and ET after the module is added. The practical focus on LMIC data with low-field scanners and inconsistent protocols is the clearest strength; pre-training to handle the domain gap is a sensible choice for a small, noisy dataset. The fusion itself is also a reasonable engineering step when single models may not be enough. What the work does not show is any evidence that the topology module is responsible for the gains. The abstract states the scores are better with the module but gives no corresponding numbers without it, no ablation table, no statistical tests, and no error analysis. The stress-test concern holds: without those controls it is impossible to separate the module's contribution from the pre-training or the model fusion. The paper also does not quantify scanner or protocol differences between the two datasets or test whether the topology step introduces new inaccuracies. These gaps make the central claim hard to evaluate from the given text. The full manuscript might contain the missing tables and implementation details, but on the evidence here the topology addition remains an unverified tweak. This kind of work is useful for teams already running BraTS-style challenges or building segmentation tools for resource-limited settings. Readers looking for new theoretical ideas or first-principles methods will find little. I would bring it to a reading group to discuss the dataset challenges and practical transfer issues, but not for the modeling novelty. I would not cite it in my own work without the ablations. It deserves peer review because the target data matters and the pipeline is reproducible in principle; referees can reasonably ask for the missing comparisons and still find value in the application.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes fusing nnU-Net and MedNeXt for brain tumor segmentation on the low-quality BraTS-Africa dataset. It pre-trains the models on BraTS 2025 adult glioma data, fine-tunes on BraTS-Africa, and adds a topology refinement module to correct prediction deformations arising from topological errors. The central empirical claim is that this module yields improved Normalized Surface Distance (NSD) scores of 0.810 (SNFH), 0.829 (NETC), and 0.895 (ET).

Significance. If the reported NSD gains can be rigorously isolated to the topology module via ablations and baselines, the approach could offer a practical route to more reliable segmentation on heterogeneous, low-field MRI data typical of LMIC settings. The work directly targets a real clinical gap in Sub-Saharan Africa, but its current evidentiary gaps prevent assessment of whether the topology component delivers the claimed correction without introducing new errors.

major comments (3)

[Abstract] Abstract: the claim that the topology refinement module produces 'better' NSD scores of 0.810/0.829/0.895 is unsupported because no baseline NSD values (without the module, or for nnU-Net/MedNeXt alone) are reported. This omission makes it impossible to attribute any improvement to topology correction versus pre-training, fusion, or fine-tuning choices.
[Abstract] Abstract: no ablation studies, implementation details of the topology refinement module, or statistical tests (e.g., paired significance on NSD deltas) are provided to substantiate the central claim that the module corrects topological deformations without introducing new inaccuracies on the BraTS-Africa data.
[Abstract] The transfer assumption from BraTS 2025 to BraTS-Africa is stated but untested; no domain-shift metrics or cross-dataset performance comparisons are given to show that pre-training actually mitigates scanner/protocol differences.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments highlight important gaps in the presentation of our results, and we address each point below. We have revised the manuscript to provide the requested baselines, ablations, details, and comparisons where feasible.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the topology refinement module produces 'better' NSD scores of 0.810/0.829/0.895 is unsupported because no baseline NSD values (without the module, or for nnU-Net/MedNeXt alone) are reported. This omission makes it impossible to attribute any improvement to topology correction versus pre-training, fusion, or fine-tuning choices.

Authors: We agree that the abstract's phrasing of 'better' NSD scores requires supporting baselines to isolate the topology module's contribution. In the revised manuscript, we will add a results table reporting NSD for nnU-Net alone, MedNeXt alone, the fused model, and each variant with/without the topology refinement module (all under the same pre-training and fine-tuning protocol). This will allow direct attribution of gains to the topology component. revision: yes
Referee: [Abstract] Abstract: no ablation studies, implementation details of the topology refinement module, or statistical tests (e.g., paired significance on NSD deltas) are provided to substantiate the central claim that the module corrects topological deformations without introducing new inaccuracies on the BraTS-Africa data.

Authors: We acknowledge this evidentiary gap in the original submission. The methods section will be expanded with full implementation details of the topology refinement module (including the topological loss formulation and deformation correction procedure). We will also add ablation studies (e.g., full model vs. without topology module) and report statistical tests such as paired Wilcoxon signed-rank tests on per-case NSD differences to confirm that corrections do not introduce new errors. revision: yes
Referee: [Abstract] The transfer assumption from BraTS 2025 to BraTS-Africa is stated but untested; no domain-shift metrics or cross-dataset performance comparisons are given to show that pre-training actually mitigates scanner/protocol differences.

Authors: The pre-training on BraTS 2025 is motivated by the need for robust initialization on heterogeneous low-field data, but we agree explicit validation is needed. In revision, we will include cross-dataset comparisons: NSD and Dice for models trained from scratch on BraTS-Africa versus pre-trained on BraTS 2025 then fine-tuned. While we did not compute explicit domain-shift metrics (e.g., MMD) in the original work, the performance deltas will demonstrate the transfer benefit. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results from standard pre-train/fine-tune pipeline

full rationale

The paper reports measured NSD scores (0.810/0.829/0.895) as direct outputs of a conventional ML pipeline: pre-training nnU-Net/MedNeXt fusion on BraTS 2025, fine-tuning on BraTS-Africa, and post-hoc addition of a topology refinement module. No equations, parameter fits, or definitions are provided that would make the reported metrics equivalent to the inputs by construction. The central claims rest on empirical evaluation rather than any self-referential derivation, self-citation chain, or renamed ansatz. Absence of ablation tables affects evidential strength but does not create circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the unverified transfer effectiveness of pre-training and the corrective power of the newly introduced topology module, both of which lack independent evidence or detailed validation in the abstract.

axioms (2)

domain assumption Pre-training on BraTS 2025 adult glioma data transfers effectively to BraTS-Africa dataset
Invoked to justify the two-stage training process for handling low-quality data.
ad hoc to paper Topology refinement module can correct prediction deformations caused by topological errors
The paper introduces this to address observed issues in model outputs.

invented entities (1)

topology refinement module no independent evidence
purpose: To refine segmentations and fix topological errors in predictions
New component added to the pipeline without external validation mentioned.

pith-pipeline@v0.9.0 · 5568 in / 1555 out tokens · 139252 ms · 2026-05-10T07:45:26.128451+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 6 canonical work pages

[1]

synapse.org/Synapse:syn53708126/wiki/626320

Brain tumor segmentation (brats) challenges - syn53708126 - wiki,https://www. synapse.org/Synapse:syn53708126/wiki/626320
[2]

Radiol Artif Intell 7(4) (2025), https://pubs.rsna.org/doi/10.1148/ryai.240528

Adewole, M., Rudie, J., Gbadamosi, A., et al.: The brats-africa dataset: Expanding the brain tumor segmentation data to capture african populations. Radiol Artif Intell 7(4) (2025), https://pubs.rsna.org/doi/10.1148/ryai.240528

work page doi:10.1148/ryai.240528 2025
[3]

Adewole, M., Rudie, J., Gbdamosi, A., et al.: The brain tumor segmentation (brats) challenge 2023: Glioma segmentation in sub-saharan africa patient pop- ulation (brats-africa) (2023)

2023
[4]

Adhikari, B., Kulung, P., Bohaju, J., et al.: Parameter-efficient fine-tuning for improvedconvolutionalbaselineforbraintumorsegmentationinsub-saharanafrica adult glioma dataset (2024),https://github.com/CAMERA-

2024
[5]

Cureus 16(12) (2024)

Bajwa, M., Najeeb, F., Alnazzawi, H., Ayub, A., Bell, J.G., Sadiq, F.: A scoping review of pakistani healthcare simulation: Insights for lower-middle-income coun- tries. Cureus 16(12) (2024)

2024
[6]

Comput Biol Med175, 108412 (2024)

Batool, A., Byun, Y.: Brain tumor detection with integrating traditional and com- putational intelligence approaches across diverse imaging modalities - challenges and future directions. Comput Biol Med175, 108412 (2024)

2024
[7]

arXiv preprint arXiv:2412.14619 (2024)

Berger, A.H., Lux, L., Weers, A., Menten, M., Rueckert, D., Paetzold, J.C.: Pitfalls of topology-aware image segmentation. arXiv preprint arXiv:2412.14619 (2024)

work page arXiv 2024
[8]

IEEE Reviews in Biomedical Engineering 16, 225–240 (2021)

Bohlender, S., Oksuz, I., Mukhopadhyay, A.: A survey on shape-constraint deep learning for medical image segmentation. IEEE Reviews in Biomedical Engineering 16, 225–240 (2021)

2021
[9]

Sensors (Basel, Switzerland) 25(6), 1838 (2025)

Bonato, B., Nanni, L., Bertoldo, A.: Advancing precision: A comprehensive review of mri segmentation datasets from brats challenges (2012–2025). Sensors (Basel, Switzerland) 25(6), 1838 (2025)

2012
[10]

The oncologist24(12), e1371–e1380 (2019)

Brand, N.R., Qu, L.G., Chao, A., Ilbawi, A.M.: Delays and barriers to cancer care in low-and middle-income countries: a systematic review. The oncologist24(12), e1371–e1380 (2019)

2019
[11]

Journal of Clinical Oncology34(1), 14–19 (2016)

Cazap, E., Magrath, I., Kingham, T.P., Elzawawy, A.: Structural barriers to diag- nosis and treatment of cancer in low-and middle-income countries: the urgent need for scaling up. Journal of Clinical Oncology34(1), 14–19 (2016)

2016
[12]

Am Fam Physician77(10), 1423–1430 (2008),https://www.aafp.org/pubs/afp/ issues/2008/0515/p1423.html

Chandana, S., Movva, S., Arora, M., Singh, T.: Primary brain tumors in adults. Am Fam Physician77(10), 1423–1430 (2008),https://www.aafp.org/pubs/afp/ issues/2008/0515/p1423.html

2008
[13]

npj Precision Oncology9, 1–13 (2025),https://www

Dorfner, F., Patel, J., Kalpathy-Cramer, J., et al.: A review of deep learning for brain tumor analysis in mri. npj Precision Oncology9, 1–13 (2025),https://www. nature.com/articles/s41698-024-00789-2

2025
[14]

In: Oh, A., Naumann, T., Glober- son, A., Saenko, K., Hardt, M., Levine, S

Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. In: Oh, A., Naumann, T., Glober- son, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural In- formation Processing Systems. vol. 36, pp. 8186–8207. Curran Associates, Inc. (2023), https://proceedings.neurips.cc/paper_files/paper/2023/file/ ...

2023
[15]

Hashmi, S., Lugo, J., Elsayed, A., Saggurthi, D., Elseiagy, M., Nurkamal, A., et al.: Optimizing brain tumor segmentation with mednext: Brats 2024 ssa and pediatrics (2024), https://arxiv.org/abs/2411.15872v2, [cited 2025 Jul 22]

work page arXiv 2024
[16]

In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., 10 Bohara et al

Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmen- tation. In: Wallach, H., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E., 10 Bohara et al. Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 32. Curran Associates, Inc. (2019),https://proceedings.neurips.cc/paper_files/ paper/2019/file/2d95666e...

2019
[17]

In: Adv Exp Med Biol, vol

Iranmehr, A., Namvar, M., Rezaei, N., Hanaei, S.: Brain and spinal cord tumors among the life-threatening health problems: An introduction. In: Adv Exp Med Biol, vol. 1394, pp. 1–18 (2023),https://link.springer.com/chapter/10.1007/ 978-3-031-14732-6_1

2023
[18]

Isensee, F., Wald, T., Ulrich, C., et al.: nnu-net revisited: A call for rigorous vali- dation in 3d medical image segmentation (2024),https://github.com/MIC-DKFZ/ nnU-Net

2024
[19]

In: Lecture Notes in Computer Science, vol

Li, L., Ma, Q., Ouyang, C., Li, Z., Meng, Q., Zhang, W., et al.: Robust seg- mentation via topology violation detection and feature synthesis. In: Lecture Notes in Computer Science, vol. 14223, pp. 67–77. Springer, Cham (2023),https: //link.springer.com/chapter/10.1007/978-3-031-43901-8_7

work page doi:10.1007/978-3-031-43901-8_7 2023
[20]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Li, L., Wang, H., Baugh, M., Ma, Q., Zhang, W., Ouyang, C., Rueckert, D., Kainz, B.: Universal topology refinement for medical image segmentation with polynomial feature synthesis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 670–680. Springer (2024)

2024
[21]

Liu, Z., Mao, H., Wu, C., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s (2025), https://github.com/facebookresearch/ConvNeXt, [cited 2025 Jul 22]

2025
[22]

Magadza, T., Viriri, S.: Efficient nnu-net for brain tumor segmentation.https: //ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10309848 (2023)

2023
[23]

Biomedicines 11(2), 364 (2023), https://www.mdpi.com/2227-9059/11/2/364/htm

Martucci, M., Russo, R., Schimperna, F., et al.: Magnetic resonance imaging of primary adult brain tumors: State of the art and future perspectives. Biomedicines 11(2), 364 (2023), https://www.mdpi.com/2227-9059/11/2/364/htm

2023
[24]

Lancet Neurol 18(4), 376–393 (2019),https://www.thelancet.com/action/showFullText?pii= S147444221830468X

Patel, A., Fisher, J., Nichols, E., et al.: Global, regional, and national burden of brain and other cns cancer, 1990–2016: a systematic analysis. Lancet Neurol 18(4), 376–393 (2019),https://www.thelancet.com/action/showFullText?pii= S147444221830468X

1990
[25]

Sci Rep10, 1–10 (2020),https: //www.nature.com/articles/s41598-020-68011-4

Poon, M., Sudlow, C., Figueroa, J., Brennan, P.: Longer-term (≥ years) survival in patients with glioblastoma: a meta-analysis. Sci Rep10, 1–10 (2020),https: //www.nature.com/articles/s41598-020-68011-4

2020
[26]

Common limitations of image processing metrics: A picture story.arXiv preprint arXiv:2104.05642,

Reinke, A., Tizabi, M.D., Sudre, C.H., Eisenmann, M., Rädsch, T., Baumgartner, M., Acion, L., Antonelli, M., Arbel, T., Bakas, S., et al.: Common limitations of image processing metrics: A picture story. arXiv preprint arXiv:2104.05642 (2021)

work page arXiv 2021
[27]

Ren, T., Honey, E., Rebala, H., Sharma, A.: An optimization framework for pro- cessing and transfer learning. Brain Tumor Segmentation, and Cross-Modality Do- main Adaptation for Medical Image Segmentation: MICCAI Challenges, BraTS 2023 and CrossMoDA 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 12 and 8, 2024, Proceedings1466...

2023
[28]

In: Medical Image Compu ting and Computer-Assisted Intervention – MICCAI 2015

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: LNCS, vol. 9351, pp. 234–241. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24574-4_28

work page doi:10.1007/978-3-319-24574-4_28 2015
[29]

Saluja, R.: Brats-2023-metrics: Official brats 2023 lesion-wise segmentation perfor- mance metrics.https://github.com/rachitsaluja/BraTS-2023-Metrics (2023), gitHub repository

2023
[30]

Neu- rosurg Rev 44(3), 1331–1343 (2020),https://link.springer.com/article/10

Verburg, N., de Witt Hamer, P.: State-of-the-art imaging for glioma surgery. Neu- rosurg Rev 44(3), 1331–1343 (2020),https://link.springer.com/article/10. 1007/s10143-020-01337-9 Topology-Driven Fusion method for Brain Tumor Segmentation 11 Supplementary Materials Axial Image Reference Standard Segmentation nnU-Net 3D nnU-Net 2D MedNeXt Coronal Sagittal B...

2020