Deep Learning Synthesized Hepatobiliary Phase Images for Optimizing Clinical Workflow of Gd-EOB-DTPA-enhanced Liver MRI: A Multicenter Study

Motivation: Our previous feasibility study showed potential of deep learning synthesized hepatobiliary phase for clinical workflow optimization. Comprehensive clinical evaluation of our preliminary findings is warranted. Goal(s): To investigate the impact of synHBP on clinical diagnosis and identification of optimized sequence combinations for model inputs Approach: Conditional generative adversarial network models trained using four sequence combinations obtained from 2049 examinations and evaluated on internal and external datasets. Results: SynHBP achieved similar differential diagnosis accuracy as acquired hepatobiliary(P>0.05). Integrating early HBP images(5 minutes) as input achieved the best results evaluated by SSIM(0.88±0.06 and 0.82±0.08), PSNR(30.65±2.81 and 28.04±3.80), lesion-parenchyma-contrast-consistency(ICC=0.91 and 0.86) for internal and external datasets, respectively. Impact: Our study demonstrated the importance of an early HBP scan of 5 minutes for high-quality HBP synthesis. Our findings help to shorten examination times and support the optimization of scanning protocols of Gd-EOB-DTPA-enhanced liver MRI.