Integrating Attention Mechanism for Accurate Skin Cancer Diagnosis

This study introduces a single Convolutional Neural Network (CNN) model enriched with attention mechanisms for skin cancer detection. Integrating attention mechanisms enhances diagnostic accuracy by dynamically highlighting critical lesion features. Outperforming conventional CNNs, the study offers both superior performance and interpretability, providing a promising solution for precise and transparent dermatologic lesion classification. The motivation for this study was that existing CNN models for skin cancer detection perform well but lack interpretability. They can accurately classify lesions, but it's difficult to understand which features in the image led to the classification. This makes it challenging for doctors to trust the diagnosis and hinders their ability to explain it to patients. The study aims to improve upon the accuracy of existing CNN models for skin cancer detection. This is crucial for the early diagnosis and timely treatment of potentially life-threatening cancers. In our study, we used the Tranasfer learning model [Inception v3] and, at the same time, developed our cnn model with attention. Our work provided training accuracy of 93% and 88% of 88% as compared to inception model of 83% training and testing accuracy of 81%.