Early Diagnosis of Multiclass Skin Lesions Using Hybrid Models Based on Fused Features

التشخيص المبكر لآفات الجلد متعددة الفئات باستخدام نماذج هجينة على أساس الميزات المندمجة

الملخص

Skin lesions (SL) are among the most serious types of skin diseases. Melanoma is considered a serious type of skin lesion. The incidence of melanoma increases annually, which poses a health risk. Life threatening. Dermoscopy is one of the best techniques that reveals invisible internal structures and helps detect types of SL. The SL are similar in the early stages, which poses a challenge to distinguish between them by manual diagnosis. Therefore, artificial intelligence (AI) techniques address deficiencies through manual diagnosis. In this study, two strategies were developed to analyze dermoscopic images for early diagnosis of SL. The images were optimized for ISIC 2018 and the Active Contour Algorithm (ACA) was applied to extract regions of interest (ROI) and isolate them from healthy areas. The ROI was fed to two strategies separately. The first strategy received the ROI and was fed to the Gray-Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP) and Fuzzy Color Histogram (FCH) algorithms to extract features. They were combined into feature vectors. The fused features were fed to the Artificial Neural Network (ANN) and Support Vector Machine (SVM) classifiers to classify them with high accuracy. The second strategy received the ROI and fed it to the ResNet18 model to extract the deep features and classify them with great efficiency using ANN and SVM. ANN-ResNet18 achieved promising results, reaching an AUC of 84.73%, sensitivity of 87.74%, accuracy of 93.8%, precision of 82.9%, and specificity of 98.47%.