Segmentation and Classification of Melanoma Lesions in Dermoscopic Images

Abstract

Skin cancer constitutes a worldwide epidemic issue. It is often detected too late because people in poor areas cannot access professional dermatologists. This study attempts to address this issue. We suggest a computerized method that uses deep learning procedures, to find and separate skin cancer lesions early on in dermoscopic pictures. Using different datasets such as ISIC and HAM10000, the approach uses MobileNetV2 and VGG16 for classification and separate the lesion mask U-Net & ResU-Net method. A robust system for early detection, increased diagnostic accuracy, and improved screening accessibility and affordability are among the main goals. With a Dice coefficient value of (94.93%), the results show that U-Net performed exceptionally well in segmentation, while MobileNetV2 achieved excellent classification accuracy (97.84%). Prioritizing data security, speed, and accuracy, the system aims to reduce mortality through rapid diagnosis. For automated melanoma detection, this method is reliable and effective. It can be implemented as a web application to encourage selfassessment and reduce the need for expensive clinical equipment.