An Effective Breast Cancer Classification System Using Multiple Feature Extraction Techniques with Multi-scale Attention-Based Feature Fusion Model.

Breast cancer ranks as the second leading cause of death in women. Early identification of breast cancer can significantly reduce women's mortality rates. Due to the time-consuming nature of manual breast cancer diagnosis, an automated system is necessary for early cancer detection. This paper presents a novel framework that combines the most useful extracted features using feature fusion for the purpose of classifying breast cancer based on a variety of features. First, the input images were obtained from the publicly available CBIS-DDSM and MIAS databases. Then, an augmentation approach is used to increase the image count. The next stage is pre-processing, which involves the use of homomorphic filtering to improve image quality. After that, various types of features, such as handcrafted, statistical, texture, as well as deep features, were extracted from the pre-processed images using different approaches. Handcrafted features were created utilizing the Histogram of Oriented Gradients (HOG) and Local Binary Pattern (LBP) approaches. Textural features such as Contrast, Entropy, Correlation, as well as Homogeneity Energy were extracted from grey-tone spatial dependence. Statistical feature extraction is carried out using the Modified Principal Component Analysis (MPCA) technique. The Vision Transformer (ViT) is used for deep feature extraction. Finally, a multi-scale attention-based feature fusion model is introduced to combine different types of features, with softmax serving as the classifier for breast cancer classification. The experiment results for the CBIS-DDSM dataset achieved an accuracy of 99.45%. Also, the MIAS dataset attains an accuracy of 99.73% compared to existing techniques; the proposed framework demonstrates superior performance.