Evolving strategies for the development and evaluation of a computerised melanoma image analysis system

The rising incidence of melanoma has increased demand for early detection tools, which enable an objective decision and prevent the need for biopsies, the current gold standard in lesion diagnosis. In this study, we report the development of such a tool, a melanoma image analysis system (MIAS), designed to improve the standard of care and classify suspicious skin lesions as benign or malignant. In addition, our methodological approach to curate an image data-set (107 benign lesions and 94 melanomas) and subsequently divide into five balanced sets in order to create a quadratic discriminant analysis classifier, allowed us to determine the causes of weaknesses within our classifier. We identified the classifier's weaknesses by analysing for which benign images the system failed and the reasons for being so. Further analysis also determined which benign categories were more difficult to classify and identified which features were useful for classification. Our classification system resulted in average sensitivity and specificity values of 79.8 and 74.8%, respectively, with an overall average accuracy of 76.6%. Developing an objective approach to visual inspection may be augmented by a methodological approach in creating a MIAS. Interestingly, the predominant (48.1%) benign misclassifications were determined to be Reed-Spitz nevi, also suggesting that in MIAS development, certain benign categories may affect future development and evaluation strategies.