Objective To explore the value of 3D and 2D radiomics features of multiparameter MRI in predicting the different expression patterns of human epidermal growth factor receptor 2(HER-2) in breast cancer before surgery, including positive versus negative (task 1), low expression versus zero expression (task 2), low expression and positive versus zero expression (task 3). Methods The MRI images of 147 patients with pathologically confirmed Luminal breast cancer were retrospectively analyzed. The patients were randomly divided into training set and test set according to the ratio of 7〖DK〗∶3. The 3D and 2D radiomics features in and around the tumor were extracted from DWI, dynamic contrast-enhanced mask (S0) and dynamic contrast-enhanced phase 3 (S3) images. Then the models were constructed by multiple pipeline combinations of three normalization methods, two dimensionality reduction methods, four feature selection methods, and ten classifiers. The receiver operating characteristic (ROC) curve and the area under the curve (AUC) were used to evaluate the prediction performance of the models in order to select the best 3D and 2D single parameter (DWI, S0, S3) and multiparameter combination (S0+S3, S0+DWI, S3+DWI, S0+S3+DWI) models. Finally, the differences between different models were compared by Delong test. Results In task 1, the AUC of 3D and 2D models in the training set was 0.777~0.832 and 0.708~0.882, respectively, and those in the test set was 0.707~0.829 and 0.702~0.846. In task 2, the AUC of 3D and 2D models in the training set was 0.779~0.870 and 0.751~0.863, respectively, and those in the test set was 0.759~0.846, 0.728~0.829. In task 3, the AUC of 3D and 2D models in the training set was 0.781~0.891 and 0.740~0.866, respectively, and those in the test set was 0.776~0.870,0.727~0.846. There was no significant statistical difference between 3D and 2D models with the same parameters. Conclusion The multiparameter MRI-based radiomics model can better predict the expression of HER-2 in breast cancer, and the models based on intratumoral and peritumoral 3D and 2D features have the same prediction efficiency