EXPERT-LEVEL RECOGNITION OF SKIN CANCER USING ADVANCED DEEP LEARNING MODELS

Abstract

The early and accurate detection of skin cancer—including melanoma, basal cell carcinoma, and squamous cell carcinoma— significantly improves patient prognosis and reduces healthcare costs worldwide. Conventional diagnosis remains heavily reliant on expert dermatologists, whose availability is often limited and subject to inter-observer variability. Advances in artificial intelligence (AI) and deep learning have opened new opportunities for automated, high-performance medical image analysis. This study presents an advanced deep learning framework comprising an ensemble of convolutional neural networks (CNNs), specifically EfficientNet-B4, ResNet50, and DenseNet201, trained on a large and diverse dataset of dermoscopic and clinical images. Our approach applies data augmentation, color normalization, and lesioncentered preprocessing to improve generalization and reduce noise. Results indicate that the ensemble achieves dermatologist-level accuracy, with an AUROC of 0.97, sensitivity of 0.90, and specificity of 0.88 across multiple skin lesion categories. Grad-CAM visualizations provide interpretability by highlighting diagnostically relevant regions of interest, enhancing clinician trust. This research demonstrates the potential of AI-powered diagnostic tools to support clinical decisionmaking in dermatology, reducing diagnostic disparities and improving early detection outcomes.