Anatomically constrained hierarchical post-processing for pediatric brain tumor segmentation

Abstract Automated segmentation of pediatric brain tumors from multi-parametric MRI is an intermediate step for clinical workflows such as treatment planning and longitudinal volumetric monitoring, yet remains challenging due to the heterogeneous appearance and complex nested anatomy of tumor sub-regions. Our evaluation of several architectural modifications—class-decoupled networks (CDHNet) and hierarchical region-aware networks (HiRA-Net)—reveals that none consistently outperform the standard nnU-Net on the BraTS-PEDs dataset. We propose Hierarchical Post-Processing (HPP), a model-agnostic pipeline that enforces anatomical constraints specific to pediatric brain tumors through: (1) connected component analysis, (2) anatomical hierarchy enforcement, (3) volume-ratio-based label correction, and (4) morphological boundary smoothing. Applied to nnU-Net, HPP improves the macro-average Dice score from 0.634 to 0.724 (+ 9.0% points), with substantial improvements for cystic component (+ 17.8 pp) and peritumoral edema (+ 13.4 pp); the same gain is reproduced across all five default cross-validation folds (raw 0.617 ± 0.021 → HPP 0.706 ± 0.023 macro Dice; ΔMacro = + 0.089 ± 0.002 per fold; pooled n = 260 paired Wilcoxon p < 10⁻⁶), confirming that the improvement is not specific to a particular data split. HPP improves all five tested backbones, demonstrating its applicability across the convolutional backbones we evaluated. We also show that region-based training, successful for adult gliomas, underperforms on pediatric tumors due to differences in label hierarchy and class distribution.