The impact of pre-processing techniques on deep learning breast image segmentation

Abstract

Abstract Breast cancer is one of the most common forms of cancer worldwide, making breast imaging a critical area for developing and evaluating Deep Learning methods. In this study, we investigate how different pre-processing techniques influence model performance in breast image segmentation. Pre-processing is a crucial step in the Deep Learning pipeline that directly impacts model performance, yet studies on its role in medical imaging remain limited. We assess the influence of different pre-processing techniques on a U-Net segmentation model applied to two breast public imaging datasets: CBIS-DDSM and Duke-Breast-Cancer-MRI. We systematically explored commonly used methods, including pixel intensity normalization, spacing harmonization, resizing/padding, and orientation standardization. Two processing pipelines were developed: Domain Non-Specific, integrating standard practices from natural and medical image analysis, and Domain Specific, which preserves anatomical information through careful handling of breast imaging metadata. A detailed comparative analysis of each pre-processing technique was conducted to evaluate its impact on model performance. Despite challenges and limitations associated with dataset size and scope, our findings identify pre-processing strategies tailored for breast imaging that can improve segmentation accuracy and analysis. This study represents an initial step in evaluating pre-processing for medical image analysis, providing a foundation for future work. Our results highlight significant differences in a 3-way ANOVA F-test ( $$\alpha < 0.05$$ ) for U-Net segmentation outcomes, attributed to different pixel intensity normalization approaches, offering valuable insights for future research.