Background <p>Variations in voxel sizes and intensities can significantly affect the reproducibility of radiomic features, yet there has been limited exploration of how image processing methods influence this reproducibility.</p> Purpose <p>This study aims to evaluate the impact of different image processing techniques on the reproducibility of radiomic features in breast cancer patients and to determine how these methods affect the performance of radiomic models.</p> Material and methods <p>Dynamic contrast-enhanced magnetic resonance imaging data from two cohorts of breast cancer patients were retrospectively analyzed. Four distinct image processing methods were applied to the original images, and 105 radiomic features were subsequently extracted from both the original and processed images. The intraclass correlation coefficient was employed to assess the reproducibility of these features. Feature selection was performed using Relief and recursive feature elimination, and three machine learning models were constructed to classify estrogen receptor (ER) and progesterone receptor (PR) status based on both original and processed images.</p> Results <p>Shape features exhibited greater robustness in reproducibility across different institutions and image processing methods compared to first-order and textural features, which were more sensitive to variations in scanners and image processing techniques. Models incorporating image processing techniques outperformed those based solely on original images in classifying ER and PR status.</p> Conclusion <p>Image processing plays a crucial role in the reproducibility of radiomic features in breast cancer, with potential implications for improving the classification of ER and PR status.</p>