A self-supervised motion-correction approach for cardiac T1 mapping

Motivation: Deep-learning algorithm has the potential to alleviate the impaction from motion in myocardial T1 mapping. However, there is no ground truth for the training. Goal(s): The aim of this study is to develop a deep learning-based algorithm to correct motion in myocardial T1 mapping using a self-supervised manner. Approach: We proposed a deep-learning approach and trained it using synthesized reference from the input T1-weighted images, eliminating the need for ground truth. Results: Our results indicated that a self-supervised deep-learning approach could align the left-ventricle myocardium and therefore improve the T1 map quaintly and accuracy. Impact: A self-supervised deep-learning approach could automatically perform motion correction for cardiovascular magnetic resonance T1 mapping, alleviating the impaction from motion and improving the quality of pixel-wise T1 map.