A Patch-Based Deep Learning Framework With 5-B Network For Breast Cancer Multi-Classification Using Histopathological Images

Abstract

Despite the fact that convolutional networks have long been the preferred architecture for investigating breast cancer (BC), new research has revealed that transformer-based architectures perform better in specific circum stances. Currently, Vision Transformer (ViT) has shown to be the most successful transformer-based architecture in vision tasks by employing patch encoding. Drawing inspiration from the patch-based modeling, this study proposed a Patch-Based deep learning network with direct operation on patches as input and the segregation of blending of spatial and channel parameters without maintaining the same size and resolution across the structure as well as simply employing conventional convolutions to carry out the mixing phases. This study further introduced a novel 5-B Network at the end of the pointwise convolutional layer for tiny feature extraction. The 5- BNet comprises five branches that process information concurrently. The primary distinction among these branches lies in the size of their convolutional kernels. In order to capture high-level image features, the 5-BNet gradually reduces the filter size of each convolution layer in every concurrent branch. Based on the proposed model, an end-to-end training for breast cancer (BC) multi-classification using the publicly available BreakHis (Benign class and Malignant Class) is carried out. In addition, an Eight-class multi-classification set by incor porating the benign four classes and the malignant four classes to further evaluate the robustness of the proposed model in multi-classification tasks. Also, the proposed model visualized internal composition to demonstrate how it understood the different patterns of the input images is illustrated. The proposed model obtained 98.1 ± 1.0% accuracy for eight classes and 98.01 ± 1.0% for four classes on all magnifications. The experimental results show that the proposed method achieves the highest breast classification accuracy when compared to cutting-edge models.