Validation of a Skin-Lesion Image-Matching Algorithm Based on Computer Vision Technology

Melanoma incidence is increasing globally, but consistently accurate skin-lesion classification methods remain elusive. We developed a simple software system to classify potentially all types of skin lesions. In the current study, we evaluated the system's ability to identify melanomas with a diameter of 10 mm or larger.

The skin-lesion classification system is composed of a proprietary database of nearly 12,000 diagnosed skin-lesion images and a computer algorithm based on the principles of content-based image retrieval. The algorithm compares characteristics of new skin-lesion images with images in the database to identify the nearest-match diagnosis.

Nearly all classification accuracy measures for this new system exceeded 90%, with results for sensitivity of 90.4% (95% confidence interval, 85.6-93.7%), specificity of 91.5% (85.4-95.2%), positive predictive value of 94.5% (90.4-96.9%), negative predictive value of 85.5% (78.7-90.4%), and overall classification accuracy of 90.8% (87.2-93.4%).

The image-matching algorithm performed with high accuracy for the classification of larger melanomas. Furthermore, the system does not require a dermoscope or any other specialized hardware; any close-focusing camera will do. This system has the potential to be an inexpensive and accurate tool for the evaluation of skin lesions in ethnically and geographically diverse populations.