ABSTRACT
To develop the method, a training set of 62 invasive melanomas and 159 clinically atypical pigmented nonmelanomas were scored for 72 surface microscopic features. Unlike in some other methods, the nonmelanoma set included nonmelanocytic lesions, such as seborrheic keratoses, hemangiomas, and dermatofi bromas. To create the model, individual features were selected with low sensitivity (0%) for melanoma, defi ning the two “negative features” for melanoma; or high specifi city (>85%) for melanoma, defi ning the 9 “positive features.” Here, “sensitivity” equals the percentage of melanomas with a particular feature and “specifi city” the percentage of nonmelanomas without that feature, for example, blue-white veil, which has a sensitivity of 51% and specifi city of 97% for the diagnosis of melanoma, meaning that it is present in 51% of melanoma and only 3% of nonmelanoma. This diagnostic method was tested on an independent set comprising 45 invasive melanomas (median Breslow thickness <0.7 mm) and 119 atypical nonmelanomas. The model gave a sensitivity of 92% and specifi city of 71% for invasive melanoma. It should be emphasized that most of the nonmelanomas used to determine specifi city were clinically atypical, leading to the decision to perform a biopsy. The use of the model would have avoided excision of 71% of those lesions. Clearly, the true specifi city of the method in the fi eld is likely to be much greater.
