Spatiotemporal Neural Diffeomorphic Flow for Cardiac Atlas and Motion Recovery

Yiyang Xu, Esther Puyol Anton, Matthew Sinclair, Amedeo Chiribiri, Steven Niederer, Alistair Young

Keywords: Cardiovascular Magnetic Resonance, Spatiotemporal Cardiac Atlases, Motion Recovery

TL;DR: We extend Neural Diffeomorphic Flow to 4D to build a topology-preserving spatiotemporal left ventricular atlas from UK Biobank CMR data, enabling motion recovery of the full cardiac cycle from as few as one frame.

Registration Requirement: Yes

Abstract: Spatiotemporal cardiac atlases built from cardiovascular magnetic resonance (CMR) data enable population-level analysis of cardiac anatomy and motion, yet existing methods lack topological guarantees and dense inter-subject correspondences. We extend Neural Diffeomorphic Flow (NDF) to the spatiotemporal domain, constructing a population-based left ventricular atlas with dense, topology-preserving correspondences across individuals and time frames. Evaluated on UK Biobank CMR data, our method achieves 1.4 mm chamfer distance for motion recovery of the full cardiac cycle from a single end-diastolic frame.

Reproducibility: https://github.com/XESchong/MIDL2026

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 65