Abstract
Multi-center liver fibrosis magnetic resonance imaging (MRI) faces cross-scanner variation and limited annotations. We study CARE 2026 GED4 (Gd-EOB-DTPA hepatobiliary-phase) liver segmentation, for which only 30 scans have manual segmentation labels. All networks are trained from scratch without external data or pretrained weights. The final system ensembles five fold-specific 2.5D student models. Each takes five adjacent slices as input, has 504,052 trainable parameters, and uses CoordConv to provide explicit spatial coordinates. Semi-supervised training uses weighted pseudo-labels as auxiliary supervision, and inference applies deterministic 3D volumetric refinement. In five-fold out-of-fold (OOF) evaluation trained only with real labels, thresholded predictions from the selected architecture achieve 0.9083 Dice and 23.67 mm HD95, improving to 0.9242 Dice and 11.01 mm HD95 after deterministic 3D volumetric refinement; most of the gain comes from retaining the largest 3D connected component. In the matched five-slice ablation, CoordConv improves Dice by 0.0206. The final Docker submission system achieves 0.9140 Dice and 35.65 mm HD95 on the official in-distribution test (Vendors A/B1/B2), and 0.9247 Dice and 21.76 mm HD95 on the Vendor C out-of-distribution test, where both images and labels are fully hidden.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CARE_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=6Ji3UCsfLy
BibTex
@InProceedings{YanZih_TopologyGuided_MICCAISAT2026,
author = { Yang, Ziheng AND Pan, Yuchi AND Liu, Ducheng AND Sun, Xue AND Xia, Qingling},
title = { { Topology-Guided Lightweight 2.5D Liver Segmentation for Limited-Annotation Multi-Center Fibrosis MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}