Purpose: To establish a predictive model for the sonication energy required for focused ultrasound surgery (FUS) of breast fibroadenomas.
Methods: This study retrospectively enrolled 87 patients with 154 benign breast tumors treated by FUS in our hospital. Radiomic analysis included 124 tumors from 69 patients, randomly split into a 3:1 ratio for training (96 cases) and validation (28 cases). Three machine learning algorithms were applied for feature selection. Then, all the selected features were used for the construction of the prediction model via four machine learning algorithms. Residual analysis and Intraclass Correlation Coefficient (ICC) analysis were performed to evaluate the performances of these four models. The importance of each feature is demonstrated by the Root Mean Square Error (RMSE) loss obtained through permutation importance measurement.
Results: This study collected 11 clinical features and 68 ultrasound radiomics features, totaling 79 independent variables. The Bagging Tree Model, characterized by lower and stable RMSE values and high R2 stability with increasing features, demonstrated superior predictive accuracy and explanatory power compared to other models. At the optimal feature count, identified by the minimum RMSE, 33 features were selected for further modeling. The bagging tree model has the highest ICC value among the four models, at 0.56, with a confidence interval of (0.23, 0.77).
Conclusions: This study established an interpretable machine learning model that integrates clinical and ultrasound radiomics features to estimate the sonication energy in FUS treatment of breast fibroadenomas.
Keywords: Focused ultrasound surgery (FUS); breast fibroadenoma; machine learning; sonication energy; ultrasound radiomics.