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

The aim of our investigation was to evaluate the usefulness of a system com posed of a digital videomicroscope equipped with a dedicated program for th e quantitative characterization of various parameters of the clinically sig nificant features of pigmented skin lesion (PSL) images, forming the basis for automatic differentiation of naevi and thin melanomas. In total 424 nae vi and 37 melanomas (including 23 thinner than 0.75 mm) were considered. Al l the digital images were acquired, framed and analysed using the DBDermo-M IPS program (Biomedical Engineering DeIl'Eva-Burroni), which calculates dif ferent parameters related to the geometry, the colour distribution and the internal pattern of the lesion. We also assessed the efficacy of an automat ic classifier, trained for 100% sensitivity using a subset of PSL images (5 9 naevi and 19 melanomas), on a test set including 365 naevi and 18 melanom as thinner than 0.75 mm. Significant differences between values from benign and malignant PSLs were observed for most of the numerical parameters. Val ues from the training set underwent elaboration by means of multivariate di scriminant analysis, enabling the identification of variables that are Impo rtant for distinguishing between the groups in order to develop a procedure for predicting group membership for new cases (test set) in which group me mbership is undetermined. Going on the training set data, a threshold score was established, enabling each melanoma to be attributed to the right grou p. When the same threshold value was employed for discriminating between be nign and malignant lesions in the test set, all the melanomas were correctl y classified, whereas 30 out of the 365 benign lesions were attributed to t he wrong group. Thus the specificity of the system reached 92%, whereas the sensitivity was 100%. Our data suggest that elaboration of videomicroscopi c images by means of dedicated software improves diagnostic accuracy for th in melanoma. Since elaboration of an image requires only 60s using our syst em, all the parameter data are available in real time and can be immediatel y examined by the classifier, providing an instant aid to clinical diagnosi s. (C) 1999 Lippincott Williams & Wilkins.