A Lightweight Residual Inception Attention Network for Explainable Skin Cancer Classification

Skin cancer is a common and potentially fatal disease caused by the unnatural growth of skin cells. It can spread to other parts of the body, and early diagnosis significantly affects survival. However, the ability to detect skin cancer early is very difficult. Therefore, in this project, a medical image binary classification system based on deep learning is designed and implemented for distinguishing benign and malignant lesions. An innovative Residual Inception Attention Model is adopted, which uses a multi-branch residual structure and combines depthwise separable convolution, spatially separable convolution, and custom attention mechanism. Interpretability analysis was performed through Grad-CAM heatmap visualization. The training results show that the model has good training results on the test set, which proves that the model has high reliability in medical image diagnosis, can effectively distinguish benign and malignant lesions, and provides strong support for clinical diagnosis.