CEUS-Based Microvascular Invasion Predictor in HCC: Improving Prognostic Stratification Following Thermal Ablation.
Microvascular invasion (MVI) critically impacts hepatocellular carcinoma (HCC) management. We aimed to develop and validate a deep learning model integrating contrast-enhanced ultrasound (CEUS) and clinical features for assessing MVI risk preoperatively, and to explore its prognostic associations in the thermal ablation (TA) cohort.
We enrolled 688 patients with solitary HCC ≤ 5 cm undergoing CEUS before surgical resection (SR) or TA. CEUS features were extracted via a vision transformer and integrated with clinical features to build an MVI classifier for SR patients. SHAP plots visualized key variables. Recurrence-free survival (RFS) was then compared between model-stratified risk groups in the TA cohort. Cox regression analysis identified risk factors for RFS.
The model achieved AUCs of 0.89 (internal validation) and 0.81 (external validation) for MVI prediction. The decision curve analysis further confirmed the clinical utility. SHAP plots identified AFP-L3%, PIVKA-II, and tumour size as the most influential clinical features. Among the CEUS features, the most significant feature was the arterial-phase dynamic sequence. In the TA cohort, patients with model-predicted high-risk MVI had significantly lower 1-, 2- and 3-year RFS rates (75.5%, 54.9%, 39.4%) compared to those with low-risk (83.8%, 80.0%, 78.0%) (p < 0.001). High MVI risk predicted by the model (HR = 2.90; 95% CI: 1.69-4.96) and AFP-L3% (HR = 1.34; 95% CI: 1.06-1.69) were independently associated with RFS in TA patients.
The model accurately predicted MVI in surgical candidates and showed exploratory prognostic value in TA patients. As a retrospective single-centre study, it warrants prospective validation.
We enrolled 688 patients with solitary HCC ≤ 5 cm undergoing CEUS before surgical resection (SR) or TA. CEUS features were extracted via a vision transformer and integrated with clinical features to build an MVI classifier for SR patients. SHAP plots visualized key variables. Recurrence-free survival (RFS) was then compared between model-stratified risk groups in the TA cohort. Cox regression analysis identified risk factors for RFS.
The model achieved AUCs of 0.89 (internal validation) and 0.81 (external validation) for MVI prediction. The decision curve analysis further confirmed the clinical utility. SHAP plots identified AFP-L3%, PIVKA-II, and tumour size as the most influential clinical features. Among the CEUS features, the most significant feature was the arterial-phase dynamic sequence. In the TA cohort, patients with model-predicted high-risk MVI had significantly lower 1-, 2- and 3-year RFS rates (75.5%, 54.9%, 39.4%) compared to those with low-risk (83.8%, 80.0%, 78.0%) (p < 0.001). High MVI risk predicted by the model (HR = 2.90; 95% CI: 1.69-4.96) and AFP-L3% (HR = 1.34; 95% CI: 1.06-1.69) were independently associated with RFS in TA patients.
The model accurately predicted MVI in surgical candidates and showed exploratory prognostic value in TA patients. As a retrospective single-centre study, it warrants prospective validation.
Authors
Chen Chen, Zhang Zhang, Huang Huang, Xu Xu, Zhang Zhang, Zhu Zhu, Wan Wan, Kong Kong, Wang Wang
View on Pubmed