Abstract

Accurate pediatric brain tumor segmentation requires reliable delineation of small and heterogeneous subregions, including enhancing tumor (ET), non-enhancing tumor (NET), and cystic component (CC), that are often underrepresented in voxel-wise training objectives. We study three nnU-Net-based strategies for the MICCAI 2026 Brain Tumor Segmentation in Pediatrics (BraTS-PED) Task 2 under a common preprocessing, inference, and postprocessing protocol: a strong 3D full-resolution nnU-Net baseline, scale-aware deep supervision, and a hierarchy-aware output parameterization. The hierarchy-aware model factorizes the BraTS-PED label structure into whole-tumor, tumor-core, and core-subtype predictions, allowing ET, NET, and CC to compete within the tumor-core branch while preserving standard nnU-Net training and export. On the official validation panel, this model improved ET Dice from 0.532 to 0.567 and CC Dice from 0.123 to 0.218 over the baseline, at the cost of small decreases on the large composite regions. We also report training-set performance. These findings suggest that label-structure-aware output modeling is a useful consideration for robust pediatric brain tumor subregion segmentation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_PEDs_012.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=ysydJAvk71

BibTex

@InProceedings{HanHao_Hierarchical_MICCAISAT2026,
        author = { Han, Haozhou AND Wang, Qian AND Ding, Haoyan},
        title = { { Hierarchical nnU-Net Logits for Pediatric Brain Tumor Subregion Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17253},
        month = {pending},
        page = {pending}
}


back to top