Deep Learning-Based Automated Diagnosis for Breast Cancer Classification Using Mammogram Analysis
Abdulrahman Aboumadi, Somaya Ali Al-Maadeed, Haya Al-Meraikhi · 2025
Breast cancer classification from screening mammograms is a critical yet challenging task due to low contrast, subtle features, and the need for reliable automated tools. While recent deep learning methods often depend on ROI annotations, multi-view fusion, or patch-based processing, these approaches introduce complexity and reduce clinical scalability. This paper proposes a streamlined, annotation-free pipeline based on an Inception v3 model trained on full mammogram images from the CBIS-DDSM dataset. The model operates in a single-view, single-model configuration and incorporates contrast enhancement (CLAHE and negative transforms), testtime augmentation (TTA), and standard online data augmentation to boost generalization. Evaluated on a fixed test set, the model achieves an AUC of 0.8444, with accuracy, precision, recall, and F1-scores all near 80%. Comparative analysis against seven state-of-the-art methods demonstrates the competitiveness of this approach under realistic conditions. To our knowledge, this represents one of the strongest performances reported for an ROI-free, single-view model on CBIS-DDSM using standard data splits.