BMEA-ViT: Breast Cancer Classification Using Lightweight Customized Vision Transformer Architecture With Multi-Head External Attention

Breast cancer is one of the primary concerns for women’s mortality across the world. Breast cancer can be diagnosed by an expert pathologist through histopathological examination of the breast lesions. Manual analysis of histopathology images by an expert pathologist is time-consuming. Therefore, automated computer-aided diagnosis plays a significant role in breast cancer prognosis. State-of-the-art (SOTA) convolution neural networks (CNNs) have earlier been utilized in the development of such systems. However, CNNs have large inductive biases and cannot model long-term dependencies in terms of global features. In this proposed work, a customised Vision Transformer (ViT) architecture with multi-head external attention, termed as BMEA-ViT, has been developed for breast cancer diagnosis. The Transformer has much lesser inductive bias and can capture long-term dependencies using an attention mechanism. However, the multi-headed self-attention (MSA) unit used in the originally proposed ViT quadratically increases its computational complexity. Therefore, in the proposed customised ViT, the MSA unit has been replaced with the multi-head external attention (MEA) unit with linear complexity. In contrast to MSA, MEA is simply implemented using a sequence of two linear layers whose weights are shared across the whole dataset. This replacement of MSA with MEA results in better generalizability with fewer inductive biases. The proposed method achieves 95.74%, 96.96%, 98.18%, and 97.25% classification accuracy on the publicly available ‘BreakHis’ dataset at four different magnification levels: $\times 40$ , $\times 100$ , $\times 200$ , and $\times 400$ , respectively. The experimental results suggest that the proposed method outperforms other recent architectures.