Noninvasive artificial intelligence system for prediction of lenvatinib based therapeutic regimens response and outcome in unresectable hepatocellular carcinoma

Purpose: Lenvatinib represents a first-line therapeutic option for unresectable hepatocellular carcinoma (uHCC), with four established treatment regimens utilized in clinical practice. However, treatment response rates remain suboptimal, with only a subset of patients achieving meaningful clinical benefit. We hypothesized that radiomics features reflecting tumor vascularization and microenvironmental heterogeneity could enhance treatment response prediction. This multi-center retrospective-prospective hybrid study aimed to develop and validate radiomic deep learning models (VALIANT) for personalized treatment response stratification in uHCC patients. Methods: This study enrolled 435 uHCC patients (2018-2024) from four institutions with histologically or radiologically confirmed unresectable disease. Radiomic features were extracted from arterial and venous phases of contrast-enhanced CT imaging acquired prior to treatment initiation. Eight VALIANT models were constructed using transfer learning strategies, with performance evaluated by area under the receiver operating characteristic curve (AUC) through five-fold cross-validation and external validation cohorts. Additionally, a prospective validation cohort of 16 patients received treatment recommendations based on the highest-scoring predicted regimen, with subsequent response assessment. Results: Partial response (PR) was achieved in 159 patients (37% response rate). The eight VALIANT models demonstrated robust performance with a mean AUC of 0.92. The arterial phase-based TPI model exhibited superior discriminative ability (training AUC 0.95; external validation AUC 0.92) with balanced classification metrics (sensitivity 0.90, specificity 0.84, accuracy 0.95), indicating excellent generalizability across institutions. All VALIANT models significantly stratified patients into distinct response/non-response groups with divergent overall survival outcomes (log-rank P < 0.01). In the prospective cohort, 13 of 16 patients (81.3%) achieved the predicted response category, with 87.5% concordance between AI-predicted optimal regimen and actual clinical response. Conclusion: The integration of deep learning radiomics and tumor phenotypic features in VALIANT models significantly enhances treatment response prediction for lenvatinib-based therapy in uHCC. This clinically applicable AI system enables comparative assessment of four lenvatinib-based regimens, facilitating personalized therapeutic selection and potentially improving patient outcomes through precision medicine approaches.