Predictive value of 18F-FDG PET/CT radiomic fusion features combined with XGBoost model for HER2 expression status in breast cancer / 中华核医学与分子影像杂志

ABSTRACT

Objective:To evaluate the value of radiomic fusion features combined with extreme gradient boosting (XGBoost) machine learning models based on 18F-FDG PET/CT images in the prediction of human epidermal growth factor receptor 2 (HER2) expression status in breast cancer. Methods:18F-FDG PET/CT images of 210 patients with primary breast cancer (all females; age 52(46, 60) years; 95 HER2-positive and 115 HER2-negative) in Tianjin Medical University Cancer Institute and Hospital between January 2012 and December 2019 were retrospectively analyzed. About 70% of the HER2-positive and HER2-negative groups were randomly selected using Python 3.7.1 software as a training set ( n=147; 67 HER2-positive and 80 HER2-negative, age 52(46, 60) years vs 55(45, 62) years) and 30% as a test set ( n=63; 28 HER2-positive and 35 HER2-negative, age 54(43, 65) years vs 52(45, 61) years). After tumor segmentation on CT and PET images being finished, CT and PET radiomic features were extracted respectively. PET/CT fusion features (including PET/CT splicing features and PET/CT mean features) were obtained through post-processing. The support vector machine (SVM) model and XGBoost model were established. The selected features were input to predict the expression status of HER2 in primary breast cancer lesions, and the prediction efficiency of the model was evaluated by ROC curve. The Delong test was used to compare the predictive effectiveness of different models and radiomic features, and the calibration curve of the machine learning model with the highest prediction efficiency was plotted. Results:Compared with SVM model, XGBoost model had better prediction performance ( z values 2.26-3.54, P values 0.016-0.040) when four kinds of radiomic features (CT features, PET features, PET/CT splicing features and PET/CT mean features) were input. ROC curve analysis showed that PET/CT mean features with XGBoost machine learning model had the best performance in predicting the expression status of HER2, and the maximum AUC was 0.83 (95% CI 0.73-0.93), which was superior to CT features (0.75(95% CI 0.63-0.88); z=3.57, P=0.027), PET features (0.73(95% CI 0.60-0.86); z=2.64, P=0.034) and PET/CT splicing features (0.74(95% CI 0.60-0.87); z=2.49, P=0.037). Conclusion:XGBoost machine learning model based on PET/CT radiomics fusion features is expected to predict HER2 expression status in patients with breast cancer.