Abstract
Graph convolutional networks (GCNs) are widely used on anatomical meshes and graphs, but their node-to-node message passing limits higher-order anatomical representation. Topological Deep Learning (TDL) addresses this by learning on higher-order domains (nodes, edges, faces), yet has mainly been used in medical imaging for segmentation. We bridge TDL with clinical risk stratification by leveraging anatomical shape for outcome prediction. Specifically, we use simplicial complex convolutional neural networks (SCCNNs) and introduce a gradient-based projection that lifts node features to edges and faces, capturing local geometric variation on cardiac surfaces. We apply this framework to ventricular arrhythmia (VA) risk prediction from cardiac CT, where left-ventricular (LV) myocardial thinning is associated with scar and abnormal electrical activity. Our pipeline segments LV masks, computes myocardial thickness, and generates a mid-wall surface mesh; thickness is used as the node feature, with higher-order information incorporated through sum or gradient-based projection. In cross-validation, a matched one-head GraphGPS baseline achieved an AUC of 79\%. SCCNN with sum projection reached 82\% AUC, while SCCNN with gradient-based projection achieved 84\% AUC, indicating competitive accuracy with a lightweight higher-order model. Beyond discrimination, Grad-CAM analysis suggests that gradient projection can produce sharper and more clinically plausible activations than sum projection, better aligned with surviving myocardium within scarred tissue and re-entry channel candidates. Compared with GraphGPS, which mainly concentrates activations on dense scar regions, gradient-based SCCNN provides anatomically interpretable localization on the cardiac surface. These results show that higher-order simplicial learning can connect anatomical shape modeling with parameter-efficient and interpretable imaging-based VA risk stratification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/GRAIL_014.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=bC638gCG9E
BibTex
@InProceedings{NjoEva_Topological_MICCAISAT2026,
author = { Njomgue Fotso, Evariste AND Ly, Buntheng AND Cochet, Hubert AND Sermesant, Maxime},
title = { { Topological Deep Learning on Graphs for Arrhythmia Risk Prediction From Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17278},
month = {pending},
page = {pending}
}