Virtual Multi-Phase Contrast Enhanced Liver MRI Using Deep Learning for Evaluating Hepatocellular Carcinoma.

A deep learning (DL) model was developed to generate contrast-enhanced MRI (CE-MRI) at multiple enhancement phases (arterial, portal venous, transitional, and hepatobiliary phases) for detecting hepatocellular carcinoma (HCC). In total, 717 patients with HCC or other non-HCC liver diseases were included, and the DL model was trained to synthesize CE-MRI. Three radiologists were invited to assess image quality, diagnostic performance, utility for the Liver Imaging Reporting and Data System (LI-RADS), and image artifacts. The image quality of DL-synthesized CE-MRI was non-inferior to that of actual CE-MRI (p-values < 0.001), and there was excellent agreement between DL-synthesized and actual CE-MRI for visualizing the LI-RADS major features. The diagnostic performance of DL-synthesized CE-MRI for HCC (sensitivity = 0.880, specificity = 0.950, AUC = 0.915) was non-inferior to that of actual CE-MRI. Notably, the proposed DL model required only 0.20-0.60 s to obtain multi-phase CE-MRI, compared with more than 20 min needed for actual multi-phase CE-MRI. Overall, the observed advantages of the proposed DL-based strategy included eliminating the need for GBCAs, high time efficiency, comparable image quality, high diagnostic performance, validity for LI-RADS, and robustness to image artifacts, indicating potential for clinical translation that may benefit both patients and healthcare providers.