An Advanced Deep Learning Framework for Automated Skin Cancer Detection and Classification

Abstract

Skin cancer is among the fastest-growing diseases worldwide, making early and accurate diagnosis essential for effective treatment and improved patient survival. However, the visual similarity between different skin lesions and the shortage of expert dermatologists make manual diagnosis challenging. This paper presents DeepDermaNet, an intelligent deep learning framework for automated skin cancer detection and classification using dermoscopic images. The proposed system utilizes the HAM10000 dataset containing 10,015 images from seven distinct skin lesion categories. To improve diagnostic performance, several preprocessing techniques, including image balancing, DullRazor hair removal, and lesion segmentation through an encoder–decoder architecture, are applied. For lesion detection and localization, multiple YOLO-based models such as YOLOv5, YOLOv6, YOLOv7, and YOLOv8 are implemented and compared. Furthermore, deep convolutional neural network models including ResNet50, DenseNet169, DenseNet201, VGG16, InceptionV3, and Xception are employed for classification. Experimental results indicate that the Xception model achieves superior accuracy and feature extraction performance compared to other architectures. The proposed framework enhances the reliability, efficiency, and precision of automated skin cancer diagnosis, thereby supporting early clinical intervention and reducing diagnostic errors in healthcare systems.