Contrast-Enhanced CT Shell Features and Deep Learning for Predicting Early Transarterial Chemoembolization Refractoriness in Hepatocellular Carcinoma.

Purpose: The aim of this study was to develop and validate a predictive model for early refractoriness to transarterial chemoembolization (TACE)-termed early TACE refractoriness (ETR)-in patients with hepatocellular carcinoma (HCC). The model integrates contrast-enhanced CT (CECT) shell features (annular features at the tumor-liver parenchyma interface) with the Vision-Mamba (Vim) architecture, known for its efficiency in handling high-resolution medical images. Patients and Methods: This study was a two-center and retrospective study. Patients from center 1 were divided into the training set (n=254) and validation set (n=108), while patients from center 2 were used as the testing set (n=75). A joint model was constructed to predict ETR, and four Vim models without clinical features and 14 machine learning models based on clinical features were also developed for comparison. Model performance was evaluated by the accuracy, area under the curve (AUC), calibration curve, sensitivity, specificity, decision curve analysis (DCA) and Delong test. SHapley Additive exPlanations(SHAP) analysis were used to explain the predictions. Results: The combined model based on the Vim framework performs better than others. The AUC of the combined model in the training set, validation set and test set were 0.959, 0.956 and 0.942, respectively. The calibration curve and DCA verified the practicality of the combined model in clinical practice. SHAP provides a visual interpretation of the model. Conclusion: The Vim-based model integrating CECT and shell features shows promise for ETR prediction, offering a preliminary stratification tool. However, it remains a promising step rather than a definitive solution, requiring prospective validation due to the retrospective design and limited validation.