Classification of skin lesions using SVM via Deep Learning feature network

Abstract

Detection of skin disease at an initial stage is significant for effective treatment. Recently, it is notable that the most threatening type of skin cancer disease among skin malignant types is melanoma since it is significantly more prone to spread to different segments of the body if not detected and treated early. Therefore, there is a requirement for a reliable and cost-effective automatic melanoma identification frameworks which offer critical help to clinicians to recognise the skin diseases as soon as possible. Such automatic image analysis frameworks give a precise and quick assessment of the lesion. In this paper, we proposed two methods to classify skin images into melanoma, benign and dysplastic, one by utilising machine learning classifier where images are preprocessed, segmented and classified into three classes of lesions utilising Support Vector Machine (SVM) and K-nearest neighbours (KNN) meta-ensemble classifier. In continuation to the first method, we proposed another strategy utilising deep learning, where dermoscopic skin lesions are enhanced and given to Convolution Neural Networks (CNNs) architecture for feature extraction and classification is performed by utilising SVM. To implement the proposed strategies, we used 200 dermoscopic skin lesions from P dataset. The outcomes conclude that deep learning feature extraction (Accuracy-92.80%) is better than that of machine learning feature extraction and classification (Accuracy — 76 %)