Improved Skin Lesion Detection with Double Layer Concatenated DenseNet Using Transfer Learning and Attention Modules

Skin that becomes unnaturally elevated, discoloured, or behaves in another unusual way may indicate a more serious health issue, such as skin cancer. Lesions on the skin, on the other hand, deviate from the skin's typical appearance and may point to more serious health issues. Accurate localization of diseased areas is challenging due to the complexity and expense of diagnostic tests; however, early lesion detection is essential for cancer prevention and diagnosis. The innovative automated method described in this article relies on Transfer Learning (TL), and its fundamental component is the use of DenseNet models. We have fine-tuned these models to repeatedly concatenate three attention modules, effectively extracting salient data for lesion identification. Before using concatenation to generate final features, it is necessary to fine-tune and incorporate attention mechanisms into all DenseNet models. For each version of DenseNet, the outcome of carrying out this method is Double Layer Concatenated DenseNet (DLCD). Using the ISIC 2019 dataset as an example, the proposed model reaches a remarkable 91.61% accuracy. Using GradCAM to pinpoint the precise region for categorization is one way to enhance the explainability of the architecture. We improve the study's accuracy while advancing the use of TL and other augmentation approaches. The results of this research provide techniques to improve the identification and early detection of skin lesions, perhaps resulting in a reduction in the incidence of skin cancer.