Triple-Phase Sequential Fusion Network for Hepatobiliary Phase Liver MRI Synthesis

Researchers from multiple Chinese institutions have introduced TriPF-Net, a novel deep learning framework designed to synthesize hepatobiliary phase (HBP) liver MRI images without requiring the typical 20-minute post-contrast delay. Published on arXiv (2604.22904), the model leverages sequential information from three pre-HBP sequences: T1-weighted imaging (the baseline), arterial-phase (AP), and venous-phase (VP). The key innovation is its ability to adaptively integrate features from AP and VP when available, and still produce robust HBP synthesis even if one or both dynamic contrast-enhanced sequences are missing—a common issue in clinical practice.

The architecture comprises an Enhanced Region-Guided Encoder and a Dynamic Feature Unification Module, optimized with a custom Region-Guided Sequential Fusion Loss. To further improve physiological consistency, the model incorporates clinical variables such as age, sex, total bilirubin, and albumin. Tested on two-center datasets, TriPF-Net achieved strong performance: on the internal dataset, MAE of 10.65, PSNR of 23.27, and SSIM of 0.76; on external validation, 12.41, 23.11, and 0.78, respectively. These results demonstrate the model's robustness and generalizability, potentially eliminating the need for delayed HBP acquisition in hepatocellular carcinoma (HCC) imaging workflows.