Enhanced Breast Cancer Classification with Neighborhood Attention Transformer and Wavelet Scattering Transform

Breast cancer is one of the most common cancers among women, occurring when abnormal cells in the breast tissue begin to grow and divide uncontrollably. While primarily affecting women, breast cancer can also develop in men, although the likelihood is much lower; less than 1% of all cases occur in men. For both men and women, however, early detection is critical for effective treatment. Early diagnosis through mammography and other imaging methods plays a crucial role in improving outcomes. In recent years, extensive studies have explored using computer systems to assist in diagnosing breast cancer, especially with artificial intelligence and computer vision applications. In this research, we propose a novel model for cancer classification in mammographic images. Our model leverages a framework capable of extracting rich and meaningful features to achieve optimal classification. Mammographic images, generated using low-energy X-rays, are first processed with the Wavelet Scattering Transform to extract precise feature maps, effectively capturing features in both frequency and spatial domains. These feature maps serve as input for the Neighborhood Attention Transformer model, which uses neighborhood attention in place of traditional selfattention. Finally, the extracted feature vectors from the mammographic images are fed into multi-layer perceptron networks to perform the classification. Notably, the proposed model has achieved a Kappa score of 40.3%, an Area Under the Curve of 76.4%, and an F1 score of 71% on the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM).