[edit]
HeartMAE: Advancing Cardiac MRI Analysis through Optical Flow Guided Masked Autoencoding
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:594-609, 2025.
Abstract
Cardiac MRI is a powerful diagnostic tool, but traditional analysis relies on complex segmentation-based workflows that may provide only a partial picture of cardiovascular health. To address these limitations, we introduce HeartMAE, a novel framework that uses masked autoencoding (MAE) to learn features directly from cardiac MRIs, without any labels. By incorporating optical flow during training, HeartMAE is guided towards cardiac regions, which significantly improves its downstream performance. A multitask model, built on a shared HeartMAE embedding layer, accurately predicts key cardiac health indicators, extracardiac features and major cardiovascular conditions. Moreover, HeartMAE features may be used as embeddings for clustering to enable patient stratification. Requiring only MRI data, HeartMAE is highly scalable and adaptable to larger datasets, paving the way for foundation models in cardiac imaging.