An Ultra-fast Deep Learning motion-compensated method for Abdominal Four-dimensional Magnetic Resonance Fingerprinting Reconstruction

Motivation: Abdominal four-dimensional (4D) magnetic resonance fingerprinting (MRF) shows great potential for benefiting image-guided radiotherapy (IGRT). However, the long reconstruction time hinders its clinical IGRT. Goal(s): This study aims to develop an ultra-fast deep learning 4D-MRF (UFDL-4DMRF) reconstruction method for benefiting clinical IGRT. Approach: The proposed UFDL-4DMRF method integrates motion compensation strategy into 4DMRF reconstruction using a three-stage deep-learning method. Twenty patients diagnosed with Hepatocellular Carcinoma were involved for training and inference. Results: UFDL-4DMRF can reduce reconstruction time by 24-fold while maintaining comparable performance in image quality, tissue property accuracy, tumor-to-tissue contrast, and motion tracking to the most advanced conventional method. Impact: UFDL-4DMRF will provide clinicians with new insights into the advantages of 4D-MRF for RT. Future work will focus on developing self-supervised deep learning models to overcome the limitations posed by the absence of ground truth data.