Standard nnU-Net brain-tumour segmentation drops 0.075 Dice on new data
single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC
What this means for you
If you validate segmentation models on held-out data from the same source as training, expect that number to fall on a new population: this pipeline lost 0.0747 Dice, concentrated in small, fragmented enhancing tumour. It is a reason to distrust single-source validation figures, not to change tools.
A conventional 3D nnU-Net — a standard, self-configuring image-segmentation pipeline — was trained on 1,351 labelled cases in five-fold cross-validation for the BraTS-GoAT 2026 brain-tumour challenge. On the official pooled validation set its Dice scores (overlap between predicted and true tumour regions; 1.0 is perfect) were 0.7805 for enhancing tumour, 0.8288 for tumour core and 0.8854 for whole tumour. Under matched fold-0 inference, mean Dice fell from 0.9058 on the source's out-of-fold cases to 0.8310 on pooled validation, a difference of 0.0747. Flipping images and averaging predictions gave small single-fold gains but no clear ensemble benefit; failures clustered in small, fragmented enhancing tumour, and a residual-encoder variant reached 0.8282.
Key facts
- ·A conventional 3D nnU-Net was trained on 1,351 labelled cases with five-fold cross-validation and 1,000 epochs per fold. source
- ·Global Dice on pooled official validation was 0.7805 for enhancing tumour, 0.8288 for tumour core and 0.8854 for whole tumour. source
- ·Under matched fold-0 inference, mean regional Dice fell from 0.9058 on source out-of-fold cases to 0.8310 on pooled validation, a difference of 0.0747. source
- ·Test-time mirroring gave small single-fold gains but no clear ensemble benefit. source
- ·A residual-encoder alternative reached 0.8282 mean Dice. source
- ·Failure cases had substantially smaller reference enhancing-tumour volumes; after adjustment for ET and WT volume, lower Dice was associated with more disconnected ET components and a smaller fraction of ET in the largest component. source
What the sources say
- Hugging Face Daily Papers (research) — Trains a standard segmentation pipeline on 1,351 cases, then measures accuracy on a new patient population.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersAssessing nnU-Net Generalization across Brain Tumor Populations in BraTS-GoAT 20262026-09-13