A deep learning model to predict objective response to TACE and TKI-based therapy in HBV-related uHCC

TACE and TKI-based combination therapy shows promise for unresectable hepatocellular carcinoma (uHCC), but inter-patient heterogeneity requires reliable biomarkers for personalized management. We developed a deep learning model to predict objective response and progression-free survival (PFS) in HBV-related uHCC. We retrospectively analyzed 243 patients, partitioned into training (clinical n = 168; radiomics n = 106) and test (n = 75) datasets. Three models were constructed: a Clinical Model (C-Model), a Machine Learning Radiomics Model (ML-Model) utilizing 1,479 CT features, and a Deep Learning Model (DL-Model) based on ResNet-50. Model interpretability was addressed via Grad-CAM. Performance was evaluated using AUC and Kaplan-Meier analysis. In the test dataset, the DL-Model achieved a superior AUC of 0.851 (95% CI: 0.747-0.954), significantly outperforming the C-Model (AUC = 0.586, P < 0.05) and exceeding the ML-Model (AUC = 0.709). Survival analysis showed the DL-Model was the only framework capable of robust prognostic stratification; predicted responders had significantly prolonged PFS (P = 0.011). Grad-CAM analysis revealed a spatial dichotomy: responders exhibited focal, centralized tumor activation, whereas non-responders showed multifocal, peripheral activation patterns. The DL-Model provides a reliable, interpretable tool for predicting tumor response and PFS in uHCC patients receiving TACE and TKI-based therapy. The Grad-CAM visualization offers spatial insights into tumor heterogeneity, facilitating personalized treatment adjustments.