Deep learning-based generation of synthetic multiphasic MRI in hepatocellular carcinoma and cirrhosis

There is growing interest in reducing contrast medium use and the lengthy scan duration in liver imaging. This proof of concept study evaluated the feasibility of deep learning-based generation of synthetic 3D liver contrast-enhanced multiphasic magnetic resonance imaging (MRI) exams, which are similar to ground-truth exams in hepatocellular carcinoma and cirrhosis.

MRI exams from patients with hepatocellular carcinoma (HCC) or cirrhosis at a single academic center were retrospectively collected. A 3D cycle-consistent generative adversarial network was trained to generate synthetic 3D T1-weighted contrast-enhanced multiphasic liver MRI exams, including arterial, portal venous, delayed, and hepatobiliary phases, using two pre-contrast T1-weighted and T2-weighted input phases. Quantitative performance evaluated similarity, error, and overlap metrics between synthetic and ground-truth exams. For the qualitative multireader study, three blinded radiologists assessed the ground-truth and synthetic MRI exams using a comprehensive questionnaire. Questionnaire tasks 1-5 comprised: visual Turing test (ground-truth vs. synthetic nature), image quality, anatomic accuracy, disease diagnosability, and artifacts. Task 6 comprised Liver Imaging Reporting and Data System features.

The study included 3,198 MRI phases from 533 MRI exams from 185 patients with HCC (mean age, 62.1 years ± 9.7 [SD]; 141 men) and 182 patients with cirrhosis (54.4 years ± 10.0; 111 men). Synthetic MRI exams achieved high quantitative and qualitative similarity to ground-truth exams. Quantitative analysis demonstrated high structural similarity index (0.86 ± 0.03), overlap (0.97 ± 0.05), and low symmetric mean absolute percent error (0.63 ± 0.23%). The qualitative multireader study showed no significant difference in tasks 1-5 (p = 0.06-0.50) and high performance metrics in task 6 (accuracy: 0.76-0.86; precision: 0.96-1.00) with moderate to perfect Fleiss's Kappa inter-rater agreement (0.58-1.00, p <0.001).

Deep learning enabled the generation of synthetic 3D liver contrast-enhanced multiphasic MRI exams from pre-contrast sequences, achieving high quantitative and qualitative similarity to ground-truth images.

This work demonstrates the early feasibility of generating high-quality, 3D contrast-enhanced multiphasic liver MRI exams from pre-contrast sequences, with synthetic exams showing strong agreement with ground-truth across quantitative metrics and key qualitative criteria, including the visual Turing test, image quality, disease diagnosability, anatomic accuracy, artifact severity, and HCC Liver Imaging Reporting and Data System features. Despite the model currently representing a proof of concept based on a moderate single-center dataset, with a need for larger multicenter studies and external validation, the results highlight the potential to transform liver MRI workflows by reducing contrast media costs and potential side effects, significantly shortening acquisition time, especially the prolonged 20-min hepatobiliary phase, and improving accessibility for patients unable to tolerate contrast-enhanced MRI because of renal impairment, contrast agent allergy, or claustrophobia.