Construction of a prediction model for chronic HBV-associated hepatocellular carcinoma based on ultrasound radiomics

A convolution neural network (CNN) prediction model was constructed based on contrast-enhanced ultrasonography (CEUS) images. The aim is to explore the predictive effect of the model for chronic HBV-associated hepatocellular carcinoma (HCC). A total of 80 cases of HBV infection in our hospital were selected retrospectively. All patients underwent CEUS, and the pathological results of the operation or liver puncture were used as the gold standard to identify HCC patients and Non-HCC patients. Two-dimensional (2D) and three-dimensional (3D) CNN prediction models are constructed based on CEUS images to predict HBV patients. Then, the prediction performance of the two models was compared. Benign and malignant lesions were confirmed in 16 and 64 patients, respectively. Compared with the 2D-CNN prediction model, the prediction accuracy of the 3D-CNN model is improved by 11.75%. The sensitivity, specificity, accuracy, and Dice coefficient of the 3D model were higher than those of the 2D prediction model. The area under the curve (AUC) of 2D and 3D models were 0.784 (95%CI: 0.671–0.897) and 0.931 (95%CI: 0.863–0.971), respectively. The 3D-CNN prediction model based on CEUS is better than the 2D model in predicting HBV-associated HCC, which is worthy of further clinical application.