Motivation: Time-resolved 4D MRI with sub-second temporal resolution is a robust technique for free-breathing imaging, while standard iterative reconstruction for time-resolved 4D MRI requires long acquisition times and high computational demand. Goal(s): This work proposes DeepGrasp, a self-supervised learning-based approach for high-quality free-breathing, time-resolved 4D MRI reconstruction with shortened scan times and reconstruction speed. Approach: DeepGrasp was developed based on self-supervised learning using an unrolled network that incorporates a low-rank subspace model-assisted training strategy and a temporal total variation constraint, enabling improved image reconstruction quality and training/inference speed. Results: DeepGrasp enables accurate 4D MRI reconstruction at high acceleration rates and fast reconstruction speed Impact: The proposed DeepGrasp technique allows for shorten data acquisition and efficient image reconstruction without requiring reference images for network training, providing significant potential for different clinical applications such as DCE-MRI or MRI-guided radiotherapy.
Pei et al. (Tue,) studied this question.