An Interpretable and Generalizable Deep Learning Model for Non-Invasive Assessment of MVI in HCC Based on Preoperative MRI: A Multi-Center Study

Motivation: MRI-based deep learning has potential for non-invasive preoperative identification of MVI in HCC. However, current models lack generalizability and interpretability. Goal(s): Develop an interpretable, domain-generalizable deep learning (DL) model for MVI assessment in HCC. Approach: A DL model using adversarial networks (AD) and different sequence combinations was validated across multiple centers. MVI-related genes were identified in a subset of HCC patients. Results: The AD-DL model showed optimal performance with an AUC of 0.793 (internal test set), 0.801 (external test set 1), and 0.773 (external test set 2). 198 MVI-related genes were identified, involving Wnt, neuroactive ligand-receptor interaction, and Hippo pathways. Impact: Incorporating adversarial networks into the model development process may enhance the generalizability of the MVI prediction model, facilitating its practical implementation.