Enhanced Skin Lesion Detection Using Concatenated DenseNet and Multi-Attention Mechanisms

Skin lesions, which often signal underlying health issues, represent deviations from normal skin appearance. Skin diseases, including the potentially serious condition of skin cancer characterized by abnormal cell growth, encompass a spectrum of health concerns. The timely detection of lesions is paramount for diagnosis and cancer prevention, yet pinpointing affected areas accurately presents challenges due to the complexity and expense of diagnostic tests. This paper presented an innovative automated solution that leveraged transfer learning through the integration of DenseNet models, fine-tuned with three attention mechanisms, to efficiently extract salient features for lesion detection. This was achieved by iteratively concatenating the models. Initially, all DenseNet models underwent incorporation and fine-tuning with attention mechanisms, followed by concatenation to derive final features. This process was reiterated across all DenseNet variants, culminating in the creation of Concatenated DenseNet (CDN). Evaluation on the ISIC 2018 dataset revealed a notable accuracy of ${9 7. 0 8 \%}$ for our proposed model. Additionally, Gradient Class Activation Map (GradCAM) was utilized to find out the specific region of interest during classification, which improved the explainability of the architecture. Our study not only enhanced accuracy but also advanced the utilization of transfer learning and various augmentation approaches. The insights from this research offered promising avenues for enhancing early detection of skin lesions, thereby contributing to the prevention of skin cancer.