Ultrasound Radiomics-Based Machine Learning and SHapley Additive exPlanations Method Predicting Pathological Prognostic Stage in Breast Cancer: A Bicentric and Validation Study.

Abstract

Purpose: To build and validate ultrasound (US) radiomics-based machine learning (ML) models to predict the pathological prognostic stage of breast cancer (BCa).

Methods: We retrospectively included 578 BCa patients from two hospitals (468 and 110 in the training and test sets, respectively). For each patient, the pathological prognostic stage was determined. US radiomics features were extracted, preprocessed, and selected. Five US radiomics-based ML models were built to distinguish between pathological prognostic stage II-III and 0-I groups. The fusion model was built by combining the optimal radiomics model with clinical indicators. Receiver operating characteristic (ROC) curve analysis was applied to evaluate the performance of the models. SHapley Additive exPlanations (SHAP) method was used to interpret the models.

Results: For each lesion, 1333 US radiomics features were extracted and 11 features were finally selected to build models. The MLP model achieved optimal performance with the AUC of 0.893 and 0.806 in the training and test sets, respectively. The AUC of the fusion model was 0.913 and 0.823 in training and test sets, respectively. The squareroot_glszm_GrayLevelNonUniformity and wavelet-LHH_gldm_DependenceVariance were the most important features of the MLP model and fusion model, respectively.

Conclusions: US radiomics-based ML is helpful for preoperatively predicting the pathological prognostic stage of BCa, which has potential reference value for making individualized treatment strategies and predicting disease prognosis in clinical practice.