SynM-FPP: Few-Shot Learning on Synthmorph-Based Motion Correction Framework for Myocardial First-Pass Perfusion

Authors

• Han Tang
• Qi Huang
• Haoteng Tang, The University of Texas Rio Grande Valley
• Liya Dai
• Caleb Berberet
• Scott Bugenhagen

Document Type

Conference Proceeding

Publication Date

3-10-2026

Abstract

First-pass perfusion (FPP) cardiac MRI is essential for assessing myocardial blood flow, yet its quantitative reliability is often limited by severe motion artifacts and rapid contrast fluctuations. We propose SynM-FPP, a two-stage, data-efficient motion correction (MoCo) framework built upon the SynthMorph model that integrates cine MRI-based pretraining, Robust Principal Component Analysis (RPCA)-derived surrogate references, and few-shot fine-tuning with segmentation-driven Dice loss. Cine pretraining provides anatomical priors. RPCA decomposition generates a rank-1 motion-free reference and identifies high-motion frames. Few-shot adaptation enables contrast-robust registration from only limited annotations. Experiments on cine and in-house FPP datasets show that our method improves temporal smoothness over existing SynthMorph baselines with significantly reduced annotation burden. This work demonstrates that combining RPCA-based motion screening with few-shot learning offers a promising path toward robust and generalizable MoCo for dynamic cardiac MRI.

Comments

©2025 IEEE

Recommended Citation

Tang, Han, Qi Huang, Haoteng Tang, Liya Dai, Caleb Berberet, Scott Bugenhagen, Thomas Schindler et al. "SynM-FPP: Few-Shot Learning on Synthmorph-Based Motion Correction Framework for Myocardial First-Pass Perfusion." In 2025 IEEE International Conference on Data Mining Workshops (ICDMW). IEEE, 2025. https://doi.org/10.1109/ICDMW69685.2025.00246

Publication Title

2025 IEEE International Conference on Data Mining Workshops (ICDMW)

DOI

10.1109/ICDMW69685.2025.00246