Abstract

By using deep learning to automate skin lesion segmentation, this work aims to improve the classification of melanoma. By properly segmenting lesions and utilizing the U-Net algorithm's preprocessing capabilities, our research aims to improve the accuracy of skin cancer diagnosis. During preprocessing, raw dermoscopic pictures from the HAM10000 dataset are enhanced and normalized early. Next, the U-Net model is used to accurately segment lesions. Advanced deep learning approaches are applied after segmentation segmented images are subjected to classification, such as Convolutional Neural Networks (CNN) and Vision Transformer (VIT) models. The VIT model demonstrated a high training accuracy of 0.94, indicating its effectiveness in learning from the training data. However, its validation and testing accuracies were at 0.73. The CNN model showed a training accuracy of 0.95, implying its ability to learn the training data effectively. However, its validation and testing accuracies were at 0.73. This all-encompassing method not only improves dermatological image analysis's dependability and effectiveness, but it also shows promise for enhancing clinical outcomes in the diagnosis and management of different forms of skin cancer. Our work is a significant step toward the creation of more reliable techniques in this important area, opening the door for improvements in patient care and healthcare diagnostics.