Skin cancer is a serious health problem, and early detection plays a key role in successful treatment and improved patient health. Traditional diagnosis mainly depends on visual examination by dermatologists, which may vary between experts and lead to inconsistencies. To reduce this variation and enhance reliability, this project proposes an automated skin cancer detection system using a multi-task deep learning approach. The proposed model performs both lesion segmentation and cancer classification simultaneously within a unified framework. The system utilizes a pretrained ResNet50 network as the encoder to extract meaningful features from dermoscopic images, and an Attention U-Net decoder to accurately segment lesion regions. A classification head with Global Average Pooling and Dense layers is used to categorize skin lesions into seven different classes. The dataset used in this work is Skin Cancer: Lesions Segmentation, which is prepared using metadata and preprocessing techniques such as resizing, normalization, and binary mask conversion. Class imbalance is addressed through oversampling of rare classes and the application of class weights during training. The final system provides lesion segmentation masks and cancer type predictions with confidence, supporting dermatologists in making faster and more consistent clinical decisions.
