Noninvasive Prediction of Ki-67 Expression in Hepatocellular Carcinoma Using Machine Learning-Based Ultrasomics: A Multicenter Study.

Abstract

To investigate the ability of ultrasomics to predict Ki-67 expression in hepatocellular carcinoma (HCC). A total of 244 patients from three hospitals were retrospectively recruited (training dataset, n=168; test dataset, n=43; and validation dataset, n=33). Lesion segmentation of the ultrasound images was performed manually by two radiologists. In total, 1409 ultrasomics features were extracted. Feature selection was conducted using the intra-class correlation coefficient, variance threshold, mutual information, and recursive feature elimination plus eXtreme Gradient Boosting. The support vector machine was combined with the learning curve and grid search parameter tuning to construct the clinical, ultrasomics, and combined models. The predictive performance of the models was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity and accuracy. The ultrasomics model performed well on the training, test, and validation datasets. The AUC (95% confidence interval [CI]) for these datasets were 0.955 (0.912-0.981), 0.861 (0.721-0.947), and 0.665 (0.480-0.819), respectively. The combination of ultrasomics and clinical features significantly improved model performance on all three datasets. The AUC (95% CI), sensitivity, specificity, and accuracy were 0.986 (0.955-0.998), 0.973, 0.840, and 0.869 on the training dataset; 0.871 (0.734-0.954), 0.750, 0.829, and 0.814 on the test dataset; and 0.742 (0.560-0.878), 0.714, 0.808, and 0.788 on the validation dataset, respectively. Ultrasomics was proved to be a potential noninvasive method to predict Ki-67 expression in HCC.