Establishment and application value of a diagnostic model for early hepatic tomors using contrast-enhanced ultrasound based on machine learning

Abstract

Objective To establish a machine learning diagnosis model for early liver tumors based on contrast-enhanced ultrasound (CEUS) images and explore its clinical diagnostic efficacy. Methods A case-control trial was conducted on the clinical data of the patients with pathologically diagnosed hepatic tumors in a diameter ≤30 mm and undergoing CEUS examinations in the First Affiliated Hospital from June 2015 to June 2020.Finally, a total of 490 patients with 520 liver masses[474 malignant tumors (91.15%) and 46 benign tumors (8.85%)]were enrolled, including 406 males (82.86%) and 84 females (17.14%), with a mean age of 51.98±0.46(22~82) years.Four sonographers were invited to analyze the features of their conventional ultrasonographs and CEUS images, and a consensus needed to be reached through discussion when opinions differ.All enrolled patients were divided into a training set (n=400) and a testing set (n=90) in a ratio of 4:1.Conventional ultrasonographs, CEUS images, history of chronic liver disease and clinical indicators were jointly analyzed using machine learning methods, such as, support vector machines, random forest, nearest neighbor algorithm, and logistic regression.Accuracy, specificity, sensitivity, and area under the receiver operating characteristic curve (AUC) were calculated to evaluate the performance of the models. Results Among the 4 models, the random forest performed best, with an accuracy, sensitivity, specificity, and AUC of 0.97, 0.83, 0.71, and 0.987(95%CI: 0.934~1.000), respectively, in the test set.There was no significant difference in the AUC values in the random forest with the other 3 models (all P>0.05).The top 5 contributing features in random forest model were cirrhosis, age, diameters in ultrasound and CEUS, and portal phase enhancement pattern, which were consistent with the clinical diagnostic indicators of malignant liver tumors reported in previous studies, indicating that our established model had a good interpretability. Conclusion Our machine learning model is established based on CEUS features, chronic liver disease and tumor markers, and shows high accuracy in diagnosis of liver tumors in diameter ≤30 mm in CEUS.

Keywords

• contrast-enhanced ultrasound
• focal liver lesion
• hepatocellular carcinoma
• machine learning