Deep Learning-Based Joint Motion Correction, Reconstruction, and Segmentation for Free-Breathing Cardiac T1 Mapping

Motivation: Motion Correction (MoCo) of free-breathing cardiac T1 mapping is imperative yet challenging due to the variable image contrast and large respiratory motion. Goal(s): To propose a deep learning-based joint motion correction, reconstruction, and segmentation method for free-breathing cardiac T1 mapping. Approach: The method exploits the inter-task dependency via physical and anatomical constraints, and employs a Coarse-to-fine Diffeomorphic Groupwise Registration Network (CDGR-Net) for motion correction. We evaluated the method on a public cardiac T1 mapping dataset. Results: Our method outperformed alternative MoCo methods in registration accuracy and mapping quality, and provided accurate segmentation of the motion-corrected T1 maps. Impact: The proposed method is expected to promote the clinical usability and reliability of cardiac T1 mapping, especially for patients who cannot tolerate repeated breath-holding MRI scans.