Enhanced Skin Lesion Classification Using Fine-Tuned DenseNet201 with Self-Attention Mechanisms

Skin lesions, if diagnosed in early stages can be treated and cured properly. Manual screening of images to identify skin lesions is time-intensive and highly prone to human errors. It also needs experts in the field to accurately diagnose the problem which is often not possible in remote areas, thereby, generating a need for automated skin lesion classification using Deep learning techniques. Hence, the paper focuses on skin lesion classification using deep learning across eight classes, employing various models. The work proposed in this paper employs the DenseNet201 model due to its superior accuracy and further fine-tunes and enhances it by integrating self-attention layers between the dense layers. This modification notably improved classification accuracy. By employing self-attention mechanisms, DenseNet201 achieved improved accuracy, highlighting its efficacy in skin lesion classification and outperforming other state-of-the-art methods such as MobileNetV2, DenseNet201, VGG-16, etc with an F-score of 80%. The study underscores the significance of model selection and architectural modifications in enhancing classification performance, particularly in medical imaging applications.