Automatic detection of critical dermoscopy features for malignant melanoma diagnosis

A method for automatically identifying a border between a skin lesion and surrounding skin on a digital image of a skin lesion comprising:

• a. preprocessing the image to identify pixels that represent features other than skin and lesion;
• b. identifying a bounding box in the shape of a rectangle having four sides by bounding box methods wherein the box borders are determined by a maxima of subtraction curves applied to horizontal and vertical projections of numbers representing a single-plane image of the lesion to determine a first lesion ratio estimate of the image;
• and wherein the absence of a bounding box side indicates that the lesion may extend to the periphery of the image, and the presence of a bounding box indicates that the lesion does not extend to the periphery of the image;
• c. if one or more bounding box sides are not found, determining a first estimate of the lesion ratio, which is the ratio of the area of the lesion to the area of the entire image, using a mean lesion ratio estimate wherein the mean lesion ratio estimate is applied to provide approximate location(s) of any missing bounding box sides;
• d. subsequently applying a lesion ratio offset, which is a correction factor, to the lesion ratio estimate to produce a corrected lesion ratio estimate, said lesion ratio offset being determined by running an automatic, self-training classifier algorithm that incorporates color and histogram data from the area defined by the bounding box sides identified in step b or c and/or by an area identified by inputting the lesion ratio estimate into a segmentation algorithm for identifying a border between a skin lesion and the surrounding skin;
• and
• e. inputting said corrected lesion ratio estimate into a segmentation algorithm for identifying a border between a skin lesion and surrounding skin;
• whereby said border is determined automatically.