Anatomy-preserving rare-subregion correction for pediatric brain tumor segmentation on multiparametric MRI: a BraTS-PEDs 2026 challenge study

A constrained rare-subregion correction within NETC improved pediatric BraTS-PEDs 2026 Task 2 cystic component segmentation Dice from 0.223 to 0.253 while preserving tumor-core and whole-tumor scores exactly.

• Why it matters: Pediatric subregion labels are imbalanced because ET, CC, and ED can be absent or tiny, creating failure modes that standard anatomy-focused models may not correct.
• What they did: The method uses a five-fold residual-encoder nnU-Net for anatomy and calibrated out-of-fold priors, then applies a rare-aware residual adapter only in NETC using MRI, signed inter-modality contrasts, uncertainty, spatial anatomy, and coordinates.
• The result: On 236 out-of-fold cases it increased CC normalized surface Dice from 0.233 to 0.263 and kept tumor-core and whole-tumor unchanged; the frozen Docker submission scored ET 0.510, NETC 0.908, CC 0.144, ED 0.000, tumor core 0.937, and whole tumor 0.937 on 91 validation cases.

AI summary of the abstract · may contain errors · not medical advice · read the paper.