Deep learning model using contrast-enhanced CT imaging features and clinical variables to predict surgical risk after hepatic resection for hepatocellular carcinoma

Background Accurate preoperative risk assessment is crucial for patients undergoing hepatic resection for hepatocellular carcinoma (HCC). Conventional prediction methods are limited by the inability to capture complex relationships among multiple factors, and existing deep learning (DL) approaches focus on individual surgical outcomes. In this study, we developed and evaluated a DL model that integrates preoperative contrast-enhanced computed tomography (CT) imaging features and clinical variables to predict intraoperative blood loss and major postoperative complications. Methods We analyzed data of 622 patients who underwent initial hepatectomy for solitary HCC between 2006 and 2021. Patients were randomly assigned to training (n = 498), validation (n = 62), and test (n = 62) cohorts. Combining convolutional neural networks and multilayer perceptron architectures, the DL model evaluated both intraoperative blood loss and postoperative complication risks. Patients were classified into major and minor intraoperative blood loss groups. Postoperative complications were defined as events of Clavien–Dindo grade IIIa or higher. Model performance was assessed via the area under the receiver operating characteristic curve (AUC). Results For predicting major intraoperative blood loss, the model achieved AUCs of 0.81 and 0.80 in the validation and test cohorts, respectively. Furthermore, predicted blood loss significantly correlated with actual blood loss (Spearman’s ρ = 0.500, p Conclusions The integrated DL model enabled the simultaneous prediction of intraoperative blood loss and postoperative complications. This framework may contribute to comprehensive preoperative surgical risk assessment, although further validation is required to confirm its clinical utility.