A cross data learning architecture for breast cancer classification using mammograms

Correct identification of breast cancer (BC) grades is of immense importance to provide targeted treatment. In this paper, a cross breast data learning (CBDL) framework for BC classification using mammogram images is developed to improve BC detection and diagnosis. In particular, a multi-step learnable approach is designed. To focus the learning model on relevant areas, we developed an image-based module to localize the breast region using adaptive thresholding combined with morphological operations. Data preprocessing and image enhancement are then applied to improve the visibility of critical mammographic features. To capture discriminatory deep visual representations of tissue features (e.g., masses, calcifications, and architectural distortion), a foundational transformer-based feature extractor model, namely BioMedCLIP, is adopted in our analysis pipeline due to its widely recognized global context awareness and multi-scale hierarchical understanding capabilities. Finally, a machine learning classifier is employed using 5-fold cross-validation. We leverage two publicly-available datasets (EMBED and VinDr-Mammo) to evaluate and generalize the model's performance. Our model achieved improved BC diagnosis (98% testing on EMBED and VinDr-Mammo) compared with other state-of-the-art work. Beyond accuracy gains, the value of this work lies in demonstrating a modular and domain-adaptive pipeline, which maintains strong performances across heterogeneous imaging environments. The combination of foundational model embeddings, feature-space harmonization, and cross-domain evaluation provides a practical path toward developing breast imaging AI systems that reliably generalize across clinical settings.