A self-supervised deep learning approach for quantitative susceptibility mapping without ground truth labels

Synopsis

Supervised deep learning methods for QSM reconstruction from tissue phase mainly rely on the ground truth susceptibility maps for training, which are not available for QSM. To address this issue, we propose a self-supervised deep learning method, TKD2TKD, for susceptibility reconstruction using artifact-contaminated TKD-TKD pairs acquired at different head orientations of one subject. The primary motivation of TKD2TKD is that the average of the network gradients by feeding TKD pairs will converge to the true gradient of that trained with high-quality averaged TKD images. The preliminary results suggested that TKD2TKD performed well in QSM reconstruction with improved susceptibility quantification accuracy.