Automatic detection of melanoma in dermoscopic images of skin lesions

Abstract

Melanoma, which is an aggressive form of skin cancer, has the highest mortality rate of all skin cancers, especially if there is a late diagnosis. The diagnosis of melanoma is usually conducted in two stages; initially an invasive biopsy of a skin lesion under concern is conducted and subsequently the respective removed tissue undergoes laboratory analysis. A crucial component of the first stage is the highly subjective process of determining which skin lesions require a biopsy. The biopsy process may be painful, costly and time consuming, however in healthcare, a case of false positive may result in a patient being unnecessarily alarmed, while a case of false negative, a type-2 error that does not detect a positive case, may have dire consequences. This is owing, in the main, to the survival rate of melanoma being correlated to the stage of the cancer when first diagnosed. This research proposes a non-invasive method to assist with the reduction of false negative classifications associated with skin lesions that are considered as benign candidates instead of melanoma. In this research, 218 dermoscopic images from the Interactive Atlas of Dermoscopy dataset were used. The selected lesions were in early stages of melanoma, being in situ or less than 0,76mm thick. Of these, 178 were used for training and 40 for testing. The training and test dataset were balanced, comprising a 50/50 split, for melanoma and non-melanoma cases. Three classification algorithms were considered in this research; these include k-nearest neighbours (KNN), Naïve Bayes and linear support vector machine (LSVM). To identify potential candidates of skin lesions for biopsy, the algorithms consider the asymmetry, border, colour, and diameter of the skin lesions; this is referred to as the ABCD rule. This research proposes the use of a LSVM machine learning algorithm to classify a skin lesion as being either melanoma or non-melanoma with a view of minimising false negative rate of the investigated classification algorithms. Classification accuracy of 87.5% and a false negative rate of 5% is achieved.