Assessing nnU-Net Generalization across Brain Tumor Populations in BraTS-GoAT 2026
Tristan Kirscher, Vivian Metzger, Philippe Meyer, Xavier Coubez
ICube, Institut Strauss / Institut Strauss / Institut Strauss, ICube
cs.CV, eess.IV
2026-09-14
Stock 3D nnU-Net scores about 0.83 mean Dice on BraTS-GoAT 2026 validation, 7.5 points below source OOF. Ensemble and residual encoder add no clear gain.
Brain-tumor segmentation has to hold up across populations. BraTS-GoAT puts adult glioma, sub-Saharan glioma, meningioma, metastases, and pediatric tumors under one protocol. Size, multiplicity, appearance, and acquisition habits all move. nnU-Net plans fitted to adult fingerprints have already underperformed on small pediatric structures.
This paper does not propose a new architecture. It freezes a conventional 3D nnU-Net, tests whether ensembling, test-time mirroring, and a residual encoder still add points, and then breaks down failures on labeled source out-of-fold predictions. Validation population labels are hidden, so every target-domain number is pooled.
Challenge data only: no external cases, no pretrained weights. The training pool is 1,351 labeled scans from adult glioma, meningioma, and metastases. Validation is 451 cases, plus sub-Saharan glioma and pediatric tumors. Inputs are T1n, T1c, T2-FLAIR, and T2w. Scoring regions are enhancing tumor (ET), tumor core (TC), and whole tumor (WT).
nnU-Net v2.6.2 in 3dfullres: 1 mm isotropic spacing, z-score inside the nonzero mask, 128x160x112 patches, batch size 2. Five folds, 1,000 epochs each, Dice plus BCE, SGD with Nesterov 0.99 and a 0.01 polynomial schedule. Each fold took about 12.3 h on one A100 40 GB. The epoch-1000 checkpoint is kept; challenge validation is not used for selection.
The submitted predictor averages the five folds and mirrored views on axes (0, 1, 2). No connected-component cleanup, no tuned region threshold. Ablations add single-fold and no-mirror variants, plus a ResidualEncoderUNet-L with larger patches. Failures are the bottom decile of strict mean regional Dice; reference-mask volume and connectivity are then correlated after volume adjustment.
The server scored 450 of 451 cases (case 02845 dropped). Five-fold plus mirroring posts global DSC 0.7805 / 0.8288 / 0.8854 for ET / TC / WT, mean DSC 0.8315 and NSD 0.5164, the best of the five submissions. The residual encoder's mean Dice is 0.8282; paired intervals against the plain U-Net include zero.
Mirroring helps a single fold a little: fold-0 ET DSC +0.0021, WT +0.0011. After five-fold averaging, every DSC and NSD change sits within 0.0008 of zero and all intervals include zero. Five folds also give no clear paired DSC gain over fold 0.
The matched comparison is sharper. The same fold-0-plus-mirror checkpoint scores 0.9058 mean Dice on source out-of-fold cases and 0.8310 on pooled validation, a gap of -0.0747 (interval -0.0976 to -0.0523). ET drops 0.0856, TC 0.0923, WT 0.0464. Hidden labels block any population-wise attribution.
Failure cases have median reference ET volume 1.04 mL against 18.66 mL for the rest. After adjusting for ET and WT volume, a larger dominant ET component still tracks higher Dice (ρ=0.350) and more ET components track lower Dice (ρ=-0.259). The ET/WT volume ratio adds nothing once size is controlled.
For contest entries and for clinical transfer, this is a negative-result baseline. A stock nnU-Net is already strong; stacking an ensemble, mirroring, and a residual encoder does not move the hidden-population leaderboard. Budget belongs on small and fragmented enhancing tumor, not on a fancier encoder.
Pooled validation near 0.83 mean Dice is not "only a small domain hit." Source out-of-fold is 0.91. The 7.5-point gap mixes evaluator details, case mix, and possible shift, and without population labels those three cannot be separated.
Validation labels and the case-to-population map are hidden, so the source-to-validation difference is not a pure shift estimate. Five-fold plus mirroring was picked post hoc on validation DSC and NSD. The server never disclosed the NSD tolerance or its unit. Failure hypotheses and estimates share the same cases; folds share training patients; intervals omit retraining uncertainty. Morphology associations on source masks do not transfer by themselves to SSA or pediatrics. No external segmentation system is compared.