Radiomics Model Based on Preoperative Contrast-enhanced CT and Clinical Features for Predicting Early Recurrence of...

This study aimed to explore the performance of models based on clinical factors combined with radiomics features on preoperative contrast-enhanced computed tomography (CECT) in predicting early postoperative recurrence of hepatocellular carcinoma (HCC).

This retrospective study included 123 patients with pathologically confirmed HCC from January 2018 to December 2021. Patients were randomly assigned to a training set (n = 96) and a testing set (n = 27). Lesions were manually contoured using ITK-SNAP, and radiomics features were extracted using PyRadiomics. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression and recursive feature elimination (RFE), followed by logistic regression (LR) to build prediction models based on clinical factors, radiomics features, or both. Model performance was assessed using the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity, and validated in the testing cohort.

The AUC of the machine learning model based on clinical factors was 0.745 (95% CI, 0.646-0.844) in the training set and 0.600 (95% CI, 0.362-0.838) in the testing set, while the AUC for the radiomics model was 0.805 (95% CI, 0.714-0.897) and 0.821 (95% CI, 0.650-0.993), respectively. Additionally, the machine learning models based on radiomics features and clinical factors demonstrated superior predictive performance, with an AUC of 0.846 (95% CI, 0.763-0.929) in the training set and 0.836 (95% CI, 0.672-1.000) in the testing set. Hosmer-Lemeshow test showed that the nomogram had the best calibration in the training set (0.206) and maintained excellent calibration in the testing set (0.222). The decision curve analysis highlighted that the radiomics model offered the highest net benefit.

Radiomics has the potential to predict early postoperative recurrence of liver cancer. A model combining clinical and radiomics features outperforms individual models built on a single type of feature in the prediction of HCC.