This study aims to assess the performance of machine learning (ML)-based contrast-enhanced CT radiomics analysis for predicating the efficacy of anti-HER2 therapy for patients with liver metastases from breast cancer.
This retrospective study analyzed 83 patients with breast cancer liver metastases. Radiomics features were extracted from arterial phase, portal venous phase, and delayed phase images, respectively. The intraclass correlation coefficient (ICC) was calculated to quantify the reproducibility of features. The training and validation sets consisted of 58 and 25 cases. Variance threshold, SelectKBest, and LASSO logistic regression model were employed for feature selection. The ML classifiers were K-nearest-neighbor algorithm (KNN), support vector machine (SVM), XGBoost, RF, LR, and DT, and the performance of classifiers was evaluated by ROC analysis.
The SVM classifier had the highest score in portal venous phase. The results were as follows: The AUC value of the poor prognosis group in validation set was 0.865, the sensitivity was 0.77, and the specificity was 0.83. The AUC value of the good prognosis group in validation set was 0.865, the sensitivity was 0.83, and the specificity was 0.77. In arterial phase, the XGBoost classifier had the highest score. The AUC value of the poor prognosis group in validation set was 0.601, the sensitivity was 0.69, and the specificity was 0.38. The AUC value of the good prognosis group in validation set was 0.601, the sensitivity was 0.38, and the specificity was 0.69. The LR classifier had the highest score in delayed phase. The AUC value of poor prognosis group in validation set was 0.628, the sensitivity was 0.62, and the specificity was 0.67. The AUC value of the good prognosis group in validation set was 0.628, the sensitivity was 0.67, and the specificity was 0.62.
Radiomics analysis represents a promising tool in predicating the efficacy of anti-HER2 therapy for patients with liver metastases from breast cancer. The ROI in portal venous phase is most suitable for predicting the efficacy of anti-HER2 therapy, and the SVM algorithm model has the best efficiency.