Assessing nnU-Net Generalization across Brain Tumor Populations in BraTS-GoAT 2026
Tristan Kirscher, Vivian Metzger, Philippe Meyer, Xavier Coubez
ICube Laboratory, University of Strasbourg · CLCC Institut Strauss
Published September 14, 2026 · on the feed September 17, 2026 · ▲ 3 · arXiv 2609.15524
The abstract
BraTS-GoAT evaluates tumor segmentation across heterogeneous populations. We trained a conventional 3D nnU-Net on 1,351 labeled cases using five-fold cross-validation and 1,000 epochs per fold. The final predictor averaged all folds and applied test-time mirroring. On pooled official validation, global DSC values were 0.7805, 0.8288, and 0.8854 for enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Under matched fold-0 inference, mean regional Dice decreased from 0.9058 on source out-of-fold (OOF) cases to 0.8310 on pooled validation (difference--0.0747). Mirroring gave small single-fold gains but no clear ensemble benefit; a residual-encoder alternative reached 0.8282 mean Dice. In labeled OOF predictions, failure cases had substantially smaller reference ET volumes; after adjustment for ET and WT volume, lower Dice remained associated with more disconnected ET components and a smaller fraction of ET contained in the largest component.
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
Assessing nnU-Net Generalization across Brain Tumor Populations in BraTS-GoAT 2026. By Tristan Kirscher and three others, from the ICube Laboratory at the University of Strasbourg and the Institut Strauss. Published on the fourteenth of September. Its code repository is public and real: twelve source files.
BraTS-GoAT evaluates brain tumour segmentation across heterogeneous populations. This entry asks how well a conventional model generalizes there. The answer has two parts. Accuracy drops when the model moves from the population it was trained on to the pooled validation populations. And the failures cluster in particular kinds of tumour: those with small enhancing regions, and those whose enhancing tumour is broken into disconnected pieces.
The model is a conventional 3D nnU-Net, trained on one thousand three hundred and fifty-one labelled cases, with five-fold cross-validation and one thousand epochs per fold. The final predictor averaged all five folds and applied mirroring at test time. On the pooled official validation set, global Dice scores were zero point seven eight zero five for enhancing tumour, zero point eight two eight eight for tumour core, and zero point eight eight five four for whole tumour.
With inference matched on a single fold, mean regional Dice fell from zero point nine zero five eight on held-out cases from the source data to zero point eight three one on pooled validation, a drop of about zero point zero seven five. Mirroring helped a single fold slightly but gave the ensemble no clear benefit.
The failure analysis rests on the labelled out-of-fold predictions. There, failure cases had substantially smaller reference volumes of enhancing tumour. After adjusting for the volumes of enhancing and whole tumour, lower Dice was still associated with more disconnected enhancing components, and with a smaller share of the enhancing tumour sitting inside the largest component. These are associations, found for one conventional architecture in one challenge.
What emerges is a standard, well-trained model that loses ground under population shift, and whose remaining errors point at tumour shape rather than tumour size alone.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
New populations, hidden labels. Start with the challenge the authors walked into.
You enter a brain tumour segmentation challenge. Training data come from adult glioma, meningioma and metastases, but validation pools those with glioma from sub-Saharan Africa and pediatric tumours, and its labels are hidden.
The road ahead
- 1 · The problem ~6 min A segmentation model trained on some tumour populations is judged on new ones with hidden labels, and stacked ensembles hide what really generalizes.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min A plain nnU-Net lost accuracy on new populations, extra tricks gave no clear gain, and failures clustered in small, fragmented enhancing tumour.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train a stock nnU-Net with five folds, ensemble them, score the hidden validation, compare against held-out source cases, then analyse failures.
- 6 · The experiment ~6 min You’ll guess: Did averaging in mirrored predictions at test time clearly improve the five-fold nnU-Net ensemble?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.