Early Detection of Non-Melanoma Skin Lesions: A ResNet50 and SVM-Based Deep Learning Approach

Abstract

This study presents a computer-aided detection method for these skin conditions by employing deep learning techniques, specifically a ResNet50-based Convolutional Neural Network (CNN), alongside a Support Vector Machine (SVM) classifier. The aim is to improve diagnostic accuracy and accessibility through image data processing and feature extraction. The main contribution of this research is the application of deep learning for automated classification of non-melanoma skin lesions, with the goal of enhancing early detection. The models were trained and evaluated using the International Skin Imaging Collaboration (ISIC) dataset, with two test scenarios to assess their performance. In Test 4, the CNN demonstrated superior results, achieving F1-scores of 44.70% for actinic keratosis, 85.25% for dermatofibroma, 78.76% for nevus, and a perfect 100.00% for vascular lesion. In comparison, the SVM model achieved lower F1-scores: 21.88% for actinic keratosis, 27.91% for dermatofibroma, 62.46% for nevus, and 70.58% for vascular lesion. The results highlight the effectiveness of deep learning, particularly CNNs, in automated dermatological diagnosis. These findings lay the groundwork for future web and mobile applications that could support early skin cancer detection and clinical decision-making.