Abstract
OBJECTIVE: This study aimed to develop and compare predictive models for hepatocellular carcinoma (HCC) differentiation using ultrasound-based radiomics and deep learning, and to evaluate the clinical utility of a combined model. METHODS: Radiomics and deep learning models were constructed from grayscale ultrasound images. A combined model integrating both approaches was developed. Model performance was assessed using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Sensitivity, specificity, accuracy, and area under the curve (AUC) were compared, and statistical significance was evaluated with the DeLong test. RESULTS: The radiomics model achieved an AUC of 0.736 (95% CI: 0.578-0.893), while the deep learning model achieved an AUC of 0.861 (95% CI: 0.75-0.972). The combined model outperformed both, with an AUC of 0.918 (95% CI: 0.836-1.0). The DeLong test indicated a significant improvement of the combined model over the radiomics model. Calibration analysis and the Hosmer-Lemeshow test showed good agreement between predictions and outcomes (p = 0.889). DCA demonstrated a higher net clinical benefit for the combined model across a range of thresholds. CONCLUSION: Integrating radiomics and deep learning enhances the predictive accuracy of ultrasound-based models for HCC differentiation, providing a promising non-invasive approach for preoperative evaluation.