Abstract
Background: Targeted combined immunotherapy (TCI) has shown certain antitumor effects in patients with unresectable hepatocellular carcinoma(uHCC), but only a subset of patients benefit. This study aims to develop a Transformer-based radiomics model to predict the objective response to combined therapy in patients with uHCC. Methods: This multicenter, retrospective study involved 264 HCC patients who underwent contrast-enhanced MRI prior to immunotherapy. The patients were divided into a training cohort(n=180) and a validation cohort(n=84). Using a multi-instance learning approach, tumor lesions in multi-sequence MRI were segmented into cross-sectional images, and features were extracted using the ResNet50 model. The Transformer model was then trained to predict the objective response rate (ORR). The prediction process was visualized using Grad-CAM and SHAP algorithms. Model performance was assessed using ROC and DCA curves, while survival analysis was conducted using Kaplan-Meier curves. Results: Among 264 patients, one achieved complete response (0.4%), 64 experienced partial response (24.2%). The ORR was 26.1% in the training group and 21.4% in the validation group. The model demonstrated high predictive accuracy, achieving a perfect area under the curve (AUC) of 1.000. Further validation using screenshot-based model inputs revealed an AUC of 0.929 (95% CI: 0.904, 0.947), confirming the model's clinical applicability. Kaplan-Meier analysis indicated that objective responders experienced better overall survival (OS) in both the training set (HR: 0.50, 95% CI: 0.27, 0.90) and the validation set (HR: 0.28, 95% CI: 0.08, 0.91). Conclusion: The deep learning framework combining ResNet50 and Transformer has proven its clinical applicability in predicting and assessing the efficacy of targeted combination immunotherapy in unresectable hepatocellular carcinoma, providing crucial guidance for clinical treatment decisions.