Abstract
Although early detection and correct diagnosis of early-stage melanoma is the most important, discrimination between melanomas and benign nevi is often difficult even by expert dermatologists, especially when these lesions are still small. We developed an Internet-based melanoma diagnostic system and opened it for the public. In present, our system was built based on 319 dermoscopy images and achieved 90.7% of sensitivity and 80.7% of specificity by cross-validation test. Our present system provides the diagnostic results by the expression "malignancy score" between 0 and 100. Considering a translating the system into a practical application, it is required that the system also provides the reason of the diagnosis with readable format for dermatologists. In this report, we investigated to the relationship between dermatologists' recognition of 15 clinical items for dermoscopy and objectively computable image parameters. We developed linear regression models for each clinical item based on 356 calculated parameters and they provided almost equivalent results to dermatologists in recognizing clinical items. Even though it is a preliminary experiment, we confirmed that the capability of computer-based method for recognizing dermoscopic patterns with a view of dermatologists.
