Abstract
Background/Aims: Predicting treatment response to immunotherapy in hepatocellular carcinoma (HCC) is essential to improve clinical outcomes with personalized treatment strategies. This study aims to develop an AI-driven prediction model using radiomic analysis from the liver and viable HCCs on pretreatment CT to differentiate responders from non-responders. Methods: HCC patients who received immunotherapy between 2016 and 2023 with pretreatment CT scans were included. Radiomic features were extracted from the whole liver and the viable HCCs on the portal venous phase CT prior to immunotherapy. Multiple machine learning models were trained for binary classification to predict treatment response, initially using liver features (Model 1), and subsequently including both liver and tumor features (Model 2). Model performance was evaluated using three-fold cross-validation. Results: Among 55 HCC patients (median age, 69; 76.4% male) who received immunotherapy, 21 (38.2%) were responders and 34 (61.8%) non-responders by mRECIST criteria. Over 5000 radiomic features were extracted from pretreatment CT scans of the liver and viable tumors, of which approximately 100 were predictive of treatment response. Model 1 (liver) achieved an average accuracy of 77%, sensitivity of 76%, and specificity of 78%. Model 2 (liver and tumor) demonstrated improved performance, with accuracy, sensitivity, and specificity of 86%, 70%, and 94%, respectively, supporting the value of combined liver-tumor radiomics in treatment response prediction. Conclusions: This pilot study developed an AI-based model using CT-derived radiomic features to predict immunotherapy response in HCC patients. The approach may offer a non-invasive strategy to support personalized treatment planning using pretreatment CT scans.