Deep Learning-Optimized GRASP-Pro Reconstruction for Highly-Accelerated DCE-MRI

Motivation: Even though GRASP-Pro has been proven to be effective in various applicatioins, it presents a significant challenge at high acceleration/frame rates as the performance of standard GRASP-Pro reconstruction can be significantly compromised under this condition. Goal(s): This work aim to develop self-supervised learning-optimized GRASP-Pro reconstruction approach to improve 4D dynamic MRI reconstruction at high acceleration rates. Approach: This approach employs self-supervised learning to first reconstruct high-quality low-resolution images to estimate a accurate temporal basis. Subsequently, self-supervised learning is applied again to reconstruct full-resolution 4D dynamic images using a low-rank subspace-assisted network training. Results: This approach demostrates improved reconstruction quality for highly-accelerated 4D DCE-MRI. Impact: The proposed self-supervised learning-optimized GRASP-Pro enables efficient and reliable 4D MRI reconstruction. This improves reconstruction quality for highly-accelerated 4D dynamic MRI, which is useful in various applications such as DCE-MRI.