Optimizing Breast Cancer Detection With Transformer Models and Canonical Correlation Analysis

Correct identification of breast cancer grades is of immense importance to provide targeted treatment. This study presents a robust transformer-based breast cancer classification system using mammograms, focusing on enhancing early detection and diagnostic accuracy. A multi-step framework is presented, incorporating advanced preprocessing techniques to address class imbalance, isolate regions of interest, and enhance tissue visibility. The feature extraction leveraged DINOv2, a transformer-based model, for its global context awareness and multi-scale hierarchical capabilities. Canonical Correlation Analysis (CCA) was applied to identify highly correlated features between tested datasets, ensuring model generalizability across diverse imaging quality. Using a machine learning classifier with a tailored MLP architecture, the system achieved 96.74% accuracy and an AUC of 0.99 on the EMBED dataset, and 97.1% accuracy and 0.98 AUC when tested on the MIAS dataset. The framework demonstrates improved reliability in classifying breast cancer as benign or malignant, outperforming existing methods in cross-dataset validation. These results underline the potential of this pipeline to enhance breast cancer diagnostics using diverse datasets, including EMBED and MIAS.