Multi-class Skin Lesion Classification Using Intelligent Techniques

This study explores the development and assessment of an advanced skin cancer classification system utilizing deep learning techniques for accurate discrimination among diverse skin cancer types. The primary aim is to construct an effective classification framework based on deep learning that has the capability of distinguishing melanoma, vascular lesions, melanocytic nevus, cutaneous fibromas, benign keratosis, and various carcinomas and skin moles. Evaluation of the system involves comprehensive analysis employing metrics like accuracy, recall, precision, and F1-score using HAM10000 dataset. We utilized pre-trained models, like EfficientNetB0, VGG19, MobileNet, and ResNet50V2 and analyzed the performance for unbalanced and manually balanced dataset for different hyperparameters. The AI-based automatic system serves as a crucial tool in aiding accurate skin cancer diagnosis. It was found that MobileNet and ResNet50V2 performed quite well as compared to other models after dataset augmentation achieving overall accuracy of 84% and average F1-score across classes of 84% for both MobileNet and ResNet50V2. While performance may vary based on image quality, dataset diversity, and population-specific factors, this research highlights the efficacy and potential of deep learning in advancing skin cancer diagnosis.