EMDA-Net: Earth Mover’s Distance (EMD) influenced Attention-aided Neural Network for Medical Image Classification

Medical image analysis methods need to be more generalized in order to meet the substantial problems associated with cancer detection and treatment, including skin, cervical, and breast cancer. In this work, we propose a novel lightweight convolutional neural network (CNN) architecture, named EMDA-Net (Earth Mover’s Distance (EMD) influenced Attention-aided Neural Network), specifically designed to meet the requirements of cancer subtype classification. It is built on the MobileNetV2 backbone and is equipped with hybrid statistical mechanism-aided attention mechanisms. Our model ensures high-performance while maintaining a lightweight design by leveraging the capabilities of the MobileNetV2 architecture and the attention mechanism. To improve similarity detection across data channels, we present a novel EMD influenced attention technique. We have examined datasets related to skin, cervical, and breast cancer for experimentation. Our model especially outperforms the state-of-the-art techniques, demonstrating its effectiveness in maximizing diagnosis precision for various cancer types across the datasets. The code of the proposed model will be made public at: https://t.ly/q5DlN