Classification Performance of Breast Tumors Using Deep Learning by a Reflected Pixel-Based Augmentation Method

Automatic classification of breast tumors from ultrasound B-Mode images using deep learning is still an exciting area of research. In a deep convolutional neural network (CNN)-based approach, augmentation plays a vital role because deep CNNs typically require large computational resources. This paper presents a reflected pixel-based novel augmentation approach for breast tumor classification using deep CNN from both spatial and multi-resolution transform domain coefficient images. The Contourlet transform domain coefficient images are experienced in this study as it has vast pyramidal and directional information. Experiments are conducted on a publicly available database (UDIAT). The proposed augmentation method provides better accuracy as compared to the augmentation processes those pre-trained neural network uses traditionally in several recently reported studies.