Motivation: Cardiac motion can lead to long acquisition times and poor T1 mapping quality. Its complexity makes correction difficult. Although deep learning methods address this problem, they usually require training data, which is rarely available. Goal(s): Efficient cardiac T1 mapping requiring only 2s acquisition-time per slice, utilizing deep learning for cardiac motion correction without the need of training data. Approach: This method combines classical iterative reconstruction methods and Zero-Shot self-supervised CNNs for motion estimation and regularization. Being subject-specific, it does not require any training dataset. Results: T1 maps from the proposed method show improved details in the structure of myocardium and papillary muscles. Impact: The proposed method represents a valid alternative to supervised deep learning quantitative reconstruction methods, by employing the advantages of deep learning techniques without the need of target or training data, which is often unavailable.
Guastini et al. (Tue,) studied this question.