Enhanced breast cancer detection in mammograms using U-MDC based semantic segmentation and deep learning models for classification

Breast cancer is one of the leading causes of cancer-related deaths among women worldwide, making early and accurate detection essential. Although mammography remains the primary imaging modality for screening, its effectiveness is often hindered by noise, artifacts, and the difficulty of identifying small tumors in dense breast tissue. Existing approaches frequently suffer from high false positive and false negative rates as well as limited generalization. This paper introduces a framework that synergistically integrates sophisticated preprocessing techniques, including noise reduction using the Center Adaptive Median Filter (CEAMF), contrast enhancement, and illumination correction, with a U-MDC-based semantic segmentation module for precise tumor delineation. A customized Convolutional Neural Network (CNN) enhanced by the Willow Catkin Optimization (WCO) algorithm serves as the baseline, while fine-tuned VGG19, ResNet152, and ResNet50 models are employed for classification. The proposed models were evaluated on the CBIS-DDSM and INbreast datasets. The best-performing ResNet50 model achieved 97.80% accuracy on CBIS-DDSM and 97.10% on INbreast, along with strong precision and recall. By effectively addressing key challenges in mammography analysis, the framework achieves encouraging results on the evaluated benchmark datasets and provides a promising basis for future external clinical validation.