Abstract
Graph Neural Networks (GNNs) have shown promise in brain disorder diagnosis but remain vulnerable to noise and multi-site heterogeneity in functional connectivity data. Existing invariant graph learning methods extract stable substructures yet largely overlook sample-dependent representation uncertainty. We propose SG-VIB, a probabilistic framework built on a causal-inspired edge disentanglement backbone. Its “Filter-then-Compress” mechanism uses Salience Gating to modulate disease-relevant graph features before a Variational Information Bottleneck maps them into a stochastic latent space. Uncertainty-Aware Reliable Contrastive Learning (U-RCL) then converts posterior dispersion into detached reliability weights during optimization. Experiments on the multi-site ABIDE I cohort and an ADNI subset show consistent improvements over representative baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_009.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=EsN6vZphgg
BibTex
@InProceedings{YuZha_SGVIB_MICCAISAT2026,
author = { Yu, Zhaolong AND Zhang, Chenghao AND Xue, Li AND Lin, Zhixin AND Huang, Yebin AND Xiao, Shengtao AND Xue, Lanyan AND Chen, Riqing AND Chen, Qiufeng},
title = { { SG-VIB: Salience-Gated Variational Information Bottleneck for Robust Invariant Brain Graph Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}
