Abstract
Brain metastasis segmentation presents persistent challenges
in accurately delineating tumor subregions with high morphological variability
and heterogeneous MRI appearance, particularly for non-enhancing
tumor core (NETC) and resection cavity (RC) – structures that exhibit
significant inter-case variability and appear in a subset of patients following
surgical intervention. Standard segmentation frameworks apply
uniform supervision across all classes, an assumption that is ill-suited
to these subregions where inter-model disagreement is concentrated at
boundaries and model confidence is unreliable[7]. In this work, we propose
a reliability-weighted supervision framework that modulates the
per-class training signal based on epistemic uncertainty estimated from
cross-validation fold ensemble disagreement. Classes with high inter-fold
disagreement within a case receive down-weighted supervision, reducing
the influence of the most ambiguous subregions on model optimization.
Built on top of nnU-Net without architectural modification, our
framework is evaluated on the BraTS 2026 metastases segmentation task.
Experimental results demonstrate consistent improvement over the vanilla
nnU-Net baseline, with the largest gains observed in NETC and RC –
precisely the subregions where inter-fold ensemble disagreement is highest.
Ablation experiments reveal that pure epistemic weighting outperforms
both aleatoric uncertainty estimation and epistemic weighting combined
with inverse class frequency correction, suggesting that in the metastasis
setting, rare subregion difficulty stems from appearance variability rather
than representation deficit – a finding that distinguishes the metastasis
from the pediatric failure mode and motivates task-specific reliability
characterization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_031.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/profile?id=%7EAjesh_Saviour_Paravila1
BibTex
@InProceedings{ParAje_Epistemic_MICCAISAT2026,
author = { Paravila, Ajesh Saviour},
title = { { Epistemic Uncertainty-Guided Supervision Weighting for Brain Metastasis Sub-region 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}
}
