Abstract
Pediatric brain tumor segmentation presents unique challenges
due to the rarity of cases, heterogeneous tumor morphology, and
significant inter-rater annotation uncertainty concentrated in clinically
critical subregions such as the cystic component (CC) and edema (ED).
Standard segmentation frameworks treat all voxels as equally informative
during training, an assumption that is particularly harmful in the
pediatric setting where certain subregions suffer from both annotation
ambiguity and severe class underrepresentation. In this work, we propose
a reliability-weighted supervision framework that adapts the training
signal at the voxel level based on epistemic uncertainty estimated from
cross-validation fold ensemble disagreement, combined with inverse class
frequency correction to address representation imbalance. Voxels belonging
to underrepresented and annotation-uncertain subregions receive
calibrated supervision weights rather than uniform loss contributions,
guiding the model to learn more effectively from reliable annotations while
reducing the influence of ambiguous label boundaries. Built on top of nnUNet
without architectural modifications, our framework is evaluated on
the BraTS 2026 pediatric tumor segmentation task. Experimental results
demonstrate consistent improvements over the vanilla nnU-Net baseline,
with the largest gains observed in CC and ED - the two subregions where
ensemble uncertainty and class frequency imbalance are most pronounced.
Ablation experiments confirm that epistemic uncertainty and frequency
correction provide complementary benefits, and that class-level scalar
weighting outperforms spatially-varying weight maps under patch-based
training.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_PEDs_017.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BraTS_PEDs_017_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=eN9H55W25R
BibTex
@InProceedings{ParAje_ReliabilityWeighted_MICCAISAT2026,
author = { Paravila, Ajesh Saviour},
title = { { Reliability-Weighted Supervision with Class Frequency Correction for Pediatric Brain Tumor 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}
}