Integrating lesser omentum adipose CT in dual-phase tumor imaging: A multi-label deep learning framework for preoperative microvascular invasion prediction and survival analysis in hepatocellular carcinoma

Background and Objective
Accurate preoperative prediction of microvascular invasion (MVI) and survival risk is essential for personalized treatment in hepatocellular carcinoma (HCC). This study aimed to develop a multi-label deep learning framework to enhance prediction performance.

Methods
We developed CGAResNet18, an end-to-end dual-branch model, using retrospective data from two centers. Computed tomography (CT) of lesser omentum adipose (LOA) was channel-wise concatenated with arterial-phase tumor CT and, separately, with venous-phase tumor CT, resulting in two fused inputs that were fed into the model. Clinical data were analyzed via univariate and multivariate logistic regression to identify three independent MVI risk factors—gender, satellite nodules, and tumor size—which were used as auxiliary labels to guide training. Patients were stratified into high-risk and low-risk groups based on the model's predictions, and overall survival (OS) analysis was conducted.

Results
The model incorporating LOA features and trained with multi-label clinical data demonstrated superior performance in MVI prediction. In the internal and external test cohort, AUCs were 0.895 (95 % CI: 0.815–0.961) and 0.842 (95 % CI: 0.747–0.930). Compared with radiologists, our model significantly reduced both false-positive and false-negative rates. The Kaplan-Meier survival analysis demonstrated that patients predicted to have MVI exhibited significantly shorter OS compared to those predicted to be MVI-absent (log-rank test, p < 0.05). For patients with MVI, surgical resection of satellite nodules may improve OS.

Conclusions
Our multi-label deep learning framework accurately predicts MVI in HCC patients and enables stratified analysis of OS, which could guide personalized treatment and improve outcomes through timely intervention.