Efficacy of support vector machine model constructed based on dual-parameter MRI radiomics in predicting the expression of human epidermal growth factor receptor-2 and hormone receptor in breast cancer patients

Objective To construct a support vector machine (SVM) model based on magnetic resonance imaging (MRI) T₂WI turbo inversion recovery magnitude (TIRM) and diffusion-weighted imaging (DWI) sequences, and evaluate its predictive performance for expression levels of human epidermal growth factor receptor-2 (HER-2) and hormone receptor (HR) in breast cancer.

Methods A total of 128 breast cancer lesions underwent breast MRI before surgery or treatment were collected, and were grouped according to immunohistochemical (IHC) method or in situ fluorescence hybridization (FISH) results. ITK-SNAP software was used to outline the three-dimensional volume region of interest (VOI) on magnetic resonance TIRM and DWI sequence images, and Pyradiomics program was introduced to extract the image omics features. After normalization of the data, a recursive feature elimination method based on support vector machine-recursive feature elimination (SVM-RFE) was used to filter the features. A total of 108 cases were divided into training group and verification group according to 8∶2 ratio by random stratified sampling method, and the other 20 cases were used as external test group. SVM machine learning classifier was used to construct the image omics model. Receiver operating characteristic (ROC) curve was used to evaluate the prediction efficiency of the model. DeLong test was used to evaluate the area under the curve (AUC) of each image omics model. SHAP algorithm was used for visual analysis, and the most contributing prediction features were screened.

Results The prediction efficiency of the combined model (training group AUC=0.94, verification group AUC=0.90) for HER-2 was higher than that of TIRM model(training group AUC=0.85, verification group AUC=0.80) and single DWI model(training group AUC=0.88, verification group AUC=0.66). The AUC of combined model in the external test group was 0.89. The feature contribution of DWI sequence obtained by SHAP algorithm was great. The image omics model based on the combination of TIRM and DWI sequence features (training group AUC=0.96, verification group AUC=0.88) and the single DWI sequence features (training group AUC=0.92, verification group AUC=0.86) was better than the model based on the single TIRM sequence features (training group AUC=0.84, verification group AUC=0.68) in HR prediction. The external test group proved that the combined model had good predictive efficiency, with an AUC of 0.90. The feature contribution of TIRM sequence obtained by SHAP algorithm was great.

Conclusion The imaging omics model constructed based on the combined features of TIRM and DWI sequences in magnetic resonance imaging has good predictive efficacy for HER-2 level, and has great potential in predicting HR expression, which can provide a basis for the formulation of personalized treatment for breast cancer patients.