Computer image analysis of pigmented skin lesions

Abstract

To assist in the distinction of melanoma from benign pigmented lesions, an imaging system was developed, comprising a frame grabber, a microcomputer, a colour video camera and flash lighting with red, green and infrared filters. Over an 18-month period, video images of 70 unselected pigmented lesions for which complete diagnostic data were available, were successfully captured using the camera. Analysis software extracted features relevant to the size, colour, shape and boundary of each lesion, and these features were correlated with clinical and histological characteristics on which standard diagnoses of skin tumours are based. For discriminant analysis based on image analysis measurements, equal probabilities were assigned to three specified diagnostic groups, namely melanoma, naevi and 'other', and four of five melanomas were correctly classified when infrared data were included. However when infrared measurements were omitted, all five melanomas were correctly classified, and the overall accuracy of classification of pigmented lesions was 71%. This system holds promise as an aid in the clinical distinction of melanoma from benign pigmented skin lesions.