Digital videomicroscopy and image analysis with automatic classification for detection of thin melanomas

Abstract

The aim of our investigation was to evaluate the usefulness of a system composed of a digital videomicroscope equipped with a dedicated program for the quantitative characterization of various parameters of the clinically significant features of pigmented skin lesion (PSL) images, forming the basis for automatic differentiation of naevi and thin melanomas. In total 424 naevi and 37 melanomas (including 23 thinner than 0.75 mm) were considered. All the digital images were acquired, framed and analysed using the DBDermo-MIPS program (Biomedical Engineering Dell'Eva-Burroni), which calculates different parameters related to the geometry, the colour distribution and the internal pattern of the lesion. We also assessed the efficacy of an automatic classifier, trained for 100% sensitivity using a subset of PSL images (59 naevi and 19 melanomas), on a test set including 365 naevi and 18 melanomas thinner than 0.75 mm. Significant differences between values from benign and malignant PSLs were observed for most of the numerical parameters. Values from the training set underwent elaboration by means of multivariate discriminant analysis, enabling the identification of variables that are important for distinguishing between the groups in order to develop a procedure for predicting group membership for new cases (test set) in which group membership is undetermined. Going on the training set data, a threshold score was established, enabling each melanoma to be attributed to the right group. When the same threshold value was employed for discriminating between benign and malignant lesions in the test set, all the melanomas were correctly classified, whereas 30 out of the 365 benign lesions were attributed to the wrong group. Thus the specificity of the system reached 92%, whereas the sensitivity was 100%. Our data suggest that elaboration of videomicroscopic images by means of dedicated software improves diagnostic accuracy for thin melanoma. Since elaboration of an image requires only 60s using our system, all the parameter data are available in real time and can be immediately examined by the classifier, providing an instant aid to clinical diagnosis.