Abstract
Ultra-low-field pediatric brain MRI enables portable neuroimaging, but automatic segmentation remains difficult because of low signal-to-noise ratio, weak anatomical contrast, artifact sensitivity, and small or thin target structures. We propose a structure-aware atlas-guided expert fusion approach for multi-structure segmentation. An nnU-Net baseline first segmented 11 brain structures, and seven refinement targets were selected based on cross-validation error profiles: the corpus callosum, caudate nuclei, hippocampi, and ventricles. Atlas-guided binary experts segmented each selected structure within target-specific crops, and their outputs were integrated into the nnU-Net prediction. On the challenge validation set, the proposed approach increased the average DSC from 0.82 ± 0.07 to 0.83 ± 0.06 and reduced HD, HD95, ASSD, and RVE from 3.46 ± 0.66 to 3.34 ± 0.62, from 1.77 ± 0.56 to 1.73 ± 0.56, from 0.49 ± 0.23 to 0.47 ± 0.22, and from 0.14 ± 0.03 to 0.10 ± 0.02, respectively. These results suggest that atlas-guided local experts can provide consistent aggregate improvements over a strong nnU-Net baseline for pediatric ultra-low-field MRI segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/LISA_025.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=9ddhgHPbhT
BibTex
@InProceedings{KimSiy_AtlasGuided_MICCAISAT2026,
author = { Kim, Siyoon AND Bae, Hyeonjun AND Kim, Jeongsik AND Choi, Jinwha},
title = { { Atlas-Guided Local Expert Fusion for Pediatric Ultra-Low-Field Brain MRI Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17257},
month = {pending},
page = {pending}
}