Breast Cancer Image Classification Using External Attention Multilayer Perceptron-Based Transformer
Shemonti Barua, Md. Saiful Islam · 2024
The prevalence of breast cancer has led to increased interest in its classification among healthcare scholars. Deep learning techniques and self-attention-based transformers have gained popularity in breast image analysis. However, self-attention’s quadratic complexity limits its capacity and interpretability, disregarding sample-to-sample association. To address this issue, a study introduces a novel attention mechanism called External Attention Multi-Layer Perceptron (EAMLP). EAMLP utilizes two external, tiny, learnable, shared retention units implemented with linear and normalization layers. By leveraging the linear complexity of external attention, this approach reduces operational complexity while considering all data point connections. The study evaluates the proposed model’s effectiveness using two publicly available datasets: Breast Histopathology Images and the BreakHis dataset. The model’s performance is assessed using different image sizes and numbers of dense layers. Additionally, a self-attention-based transformer is experimentally tested for comparison. The results demonstrate that the proposed model achieves a maximum accuracy of 95.73%.