Comparative Analysis of Skin Lesions Classification Using Machine Learning Classifiers and Lesnet-22 Architecture

One of the prevalent and lethal cancers in the world is skin cancer. Skin cancer cases have been rising in recent years, and this is expected to increase exponentially. It is curative if the skin lesions are detected early. Diagnosis is important because of the similarities in its types like melanoma (MEL), seborrheic keratosis (SK), actinic keratosis (AKEIC), basal cell carcinoma (BCC), benign keratosis (BKL), dermatofibroma (DF), melanocytic nevus (NEV), squamous cell carcinoma (SCC), and vascular lesion (VASC) . Therefore, there is a growing demand for computer-assisted recognition approaches for dermoscopic images of skin lesions. Many automated methods for diagnosing skin lesions have been proposed, but they have not yet proven to be very accurate. This study proposes a computer-assisted recognition approach to classify skin lesions. Initially, the preprocessing technique eliminates the digital artifacts from the dermoscopic images, and features are extracted based on the feature fusion approach. Moreover, fully automated detection and segmentation approaches are employed to effectively localize the skin lesions. Two methods are proposed for classification: the firstly, the analysis and comparison of different machine learning approaches is performed. The second method employs a deep learning framework-based approach. To validate the proposed methodologies, dermoscopic images from International Skin Imaging Collaboration (ISIC-2017) and ISIC-2019 datasets have been used. The skin lesions classification performance of the proposed novel LesNet-22 architecture has achieved 94% and 91% of accuracy for ISIC-2017 and ISIC-2019 respectively, which outperforms the existing classifiers.