Enhancing Cardiac Functional Assessment with Deep-Learning-based Automatic Segmentation of Free-Running 4D Whole-Heart MR Images

Motivation: Manual segmentation of 5D free-running (FR) cardiac MRI data is arduous, which hinders clinical adoption of this comprehensive imaging technique. Goal(s): To develop and validate a deep learning (DL) framework for the automatic and accurate segmentation of isotropic whole-heart images from FR acquisitions. Approach: We trained a 3D nnU-Net with a residual encoder using semi-automatically generated reference standard segmentations and evaluated the accuracy of the resulting fully automated segmentations through geometric and functional metrics. Results: DL-based segmentation closely matched semi-automatic results with high Dice similarity coefficient (>0.91), low volume errors (~5%), and excellent functional parameter agreement (ICC > 0.96), confirming accuracy. Impact: Our DL framework enables automatic, consistent segmentation of FR cardiac MRI, reducing manual workload and facilitating quantitative and comprehensive 5D assessments, supporting broader clinical and research use of advanced whole-heart imaging.