DERMATOLOGICAL DIAGNOSIS OF SKIN LESIONS USING ATTENTION MECHANISMS

Every year, both in developed and developing nations, the prevalence of skin cancer rises. Due to a lack of resources and medical knowledge, diagnosing skin lesions is more difficult in third-world nations. Unusual growths or alterations in the skin are known as skin lesions, and they can result from a number of causes, including cancer, inflammatory diseases, infections, traumas, and heredity. Malignant (cancerous) and benign (non-cancerous) skin lesions are both possible. This disease is contracted when the pigments that produce skin color become cancerous. Dermatologists find it difficult to diagnose skin cancer since the pigments of various skin conditions might look identical. This led to the goal of this work, which is to use attention mechanisms to design a system for dermatological diagnoses of skin lesions. Modern network architectures ResNet and EfficientNet, enhanced with specially designed patch-based attention heads, are the approach used to accomplish this. The HAM10000 dataset, a thorough compilation of dermatoscopic pictures of typical pigmented skin lesions, was used in the investigation. In order to improve the model's capacity to recognize minute yet crucial variations among lesion types, attention heads were created to highlight and identify important characteristics within patches of the dermatoscopic pictures. According to the experimental results, the model that accurately classifies images into different lesion types had the lowest accuracy of 72% on a dataset of over 10,000 image instances, while the model that determines whether a lesion is cancerous or non-cancerous had the highest accuracy of 98%, demonstrating its robustness and reliability for...