Integrating multiphase MRI surrogates to improve microvascular invasion detection and recurrence risk stratification in HCC.

Background & Aims: Microvascular invasion (MVI) is a key determinant of recurrence and poor outcomes in hepatocellular carcinoma (HCC), yet accurate preoperative detection remains challenging. A deep-learning model integrating multiphase MRI and clinical features was developed and externally validated to non-invasively predict MVI and stratify postoperative recurrence risk. Methods: Clinico-radiological data from 924 patients with resected HCC (2014-2023, five tertiary centers in China) were retrospectively assembled. A deep-learning model (DL-TriFusion) integrating multiphase MRI and clinical variables for MVI prediction was trained (n = 361), internally validated (n = 155), and externally validated across three centers (n = 408; 188/136/84); whether recurrence risk stratification using imaging-based surrogates is superior to pathology-confirmed MVI alone was also evaluated. Statistical analysis included classification and detection metrics. Results: DL-TriFusion achieved AUCs ranging from 0.957 to 0.959, significantly outperforming all unimodal and bimodal models ( Conclusions: DL-TriFusion enables robust preoperative prediction of MVI. It also improves postoperative stratification of early and late recurrence. Impact and implications: DL-TriFusion represents a reliable, non-invasive biomarker with strong potential for clinical translation. By leveraging routinely available MRI and clinical data, it enables accurate preoperative assessment of microvascular invasion and stratification of recurrence risk. This capability could optimize surgical planning, identify candidates for neoadjuvant or adjuvant therapy, refine surveillance strategies, and reduce unnecessary overtreatment - ultimately supporting personalized management of hepatocellular carcinoma.