nnU-Net Generalization Test on BraTS-GoAT: Dice Drops from 0.906 to 0.831 Across Populations
Tristan Kirscher · hf · 2026-09-17
A study evaluates 3D nnU-Net generalization across heterogeneous brain tumor populations in BraTS-GoAT 2026.
- Trained on 1,351 labeled cases with five-fold cross-validation (1,000 epochs per fold), ensembling all folds with test-time mirroring
- Global DSC on pooled validation: 0.7805 (ET), 0.8288 (TC), 0.8854 (WT)
- Under matched fold-0 inference, mean regional Dice drops from 0.9058 on source OOF cases to 0.8310 on pooled validation (-0.0747), revealing distribution shift
- Mirroring gives small single-fold gains with no clear ensemble benefit; a residual-encoder alternative reaches 0.8282
- Failure cases have much smaller reference ET volumes; after volume adjustment, lower Dice remains tied to more disconnected ET components
More from Research
- "Time Machine Experiments": AI-simulated minds from 1930 as a research method — iyadrahwan · 2026-09-17
- NVIDIA Launches V2D Challenge to Teach Robots From Video — YunzhuLiYZ · 2026-09-17
- Tsinghua's CERA-MoA co-evolves routing and LLM agents via RL — Tsinghua · 2026-09-17
- Streaming ASR confidently botches names and numbers: a decoder that waits 200ms — Sinver_Nightingale27 · 2026-09-17
- New online inverse optimization algorithm achieves O(d) regret, spotlights Cogentic agent framework — IgorCarron · 2026-09-17
- AndroidLife: Qwen3.8-27b runs 60 real phone tasks, fails 43% and cooks the chip to 98.2°C — East-Muffin-6472 · 2026-09-17