SEARCH

SEARCH BY CITATION

Abstract

Background  For early melanoma diagnosis, experienced dermatologists have an accuracy of 64–80% using clinical diagnostic criteria, usually the ABCD rule, while automated melanoma diagnosis systems are still considered to be experimental and serve as adjuncts to the naked-eye expert prediction. In an attempt to aid in early melanoma diagnosis, we developed an image processing program with the aim to discriminate melanoma from melanocytic nevi, establishing a mathematical model to come up with a melanoma probability.

Methods  Digital images of 132 melanocytic skin lesions (23 melanomas and 109 melanocytic nevi) were studied in features of geometry, color, and color texture. A total of 43 variables were studied for all lesions, e.g., geometry, color texture, sharpness of border, and color variables. Univariate logistic regression analysis followed by “−2 log likelihood” test and Spearman's rank correlation coefficient were used to eliminate inappropriate variables, as the presence of multicollinearity among variables could cause severe problems in any stepwise variable selection method. Initially, “−2 log likelihood” and nonparametric Spearman's rho picked five variables to be included in a multivariate model of prediction. The five-variable model was then reduced to three variables and the performance of each model was tested. The “jackknife” method was performed in order to validate the model with the three variables and its accuracy was weighed vs. the five-variable model by receiver-operating characteristics (ROC) curve plotting. It was concluded that the reduced model did not compromise discriminatory power.

Results  Not all variables contributed much to the model, therefore they were progressively eliminated and the model was finally reduced to three covariates of significance. A predictive equation was calculated, incorporating parameters of geometry, color, and color texture as independent covariates for the prediction of melanoma. The proposed model provides melanoma probability with a 60.9% sensitivity and 95.4% specificity of prediction, an overall accuracy of 89.4% (probability level 0.5), and 8% false-negative results.

Conclusions  Through a digital image processing system and the development of a mathematical model of prediction, discrimination between melanomas and melanocytic nevi seems feasible with a high rate of accuracy using multivariate logistic regression analysis. The proposed model is an alternative method to aid in early melanoma diagnosis. Expensive and sophisticated equipment is not required and it can be easily implemented in a reasonably priced portable programmable computer, in order to predict previously undiagnosed skin melanoma before histopathology results confirm diagnosis.