Machine learning-based ultrasound radiomics for predicting mutation status in hepatocellular carcinoma.

OBJECTIVES

To explore the utility of machine learning-based ultrasound radiomics for predicting gene mutation in hepatocellular carcinoma (HCC).

METHODS

154 HCC patients with 182 lesions from 2019 to 2024 were reviewed retrospectively. All lesions were randomly split into the training set ( = 129) and the test set ( = 53), and ultrasound radiomics features were extracted and selected. Extreme gradient boosting tree (XGBoost), decision tree (DT), random forest (RF), support vector machine (SVM), and logistic regression (LR) were used to construct the ultrasound radiomics models, the clinical models, and the combined models. The predictive performance of various models was evaluated by the area under the curve (AUC), accuracy, calibration curve, and decision curve analysis (DCA).

RESULTS

Among the 182 lesions, 102 were confirmed as mutant and 80 were confirmed as wild-type . The ultrasound radiomics model obtained an AUC of 0.778 and an accuracy of 0.774 in the test set. The clinical model achieved an AUC of 0.761 and an accuracy of 0.710 in the test set. Notably, integrating clinical features with ultrasound radiomics further enhanced predictive performance. The XGBoost-based combined model exhibited the highest predictive performance among all models, achieving an AUC of 0.846 and an accuracy of 0.823 in the test set. The decision curve analysis and calibration curve revealed that the XGBoost-based combined model provided the highest clinical benefit and exhibited strong predictive consistency.

CONCLUSION

Machine learning-based ultrasound radiomics signatures accurately predict gene mutations in HCC. The XGBoost-based combined model, which combined ultrasound radiomics features with clinical features, showed the best performance and represented a promising noninvasive approach for screening -mutated HCC.