Automatic registration of multiple skin lesions by use of point pattern matching

Computerized comparison of serial skin images is a potentially valuable tool for melanoma screening. In automating this process, matching or "registering" each lesion in a pair of images plays an important role in looking for clinically significant change. We have investigated three practical techniques--a point pattern correlation, a 2-point geometrical transformation, and a 3-point geometrical transformation--for their effectiveness in matching and identifying lesions in pairs of skin images. These techniques view the spots in each image as a point pattern to be matched from image to image. Each of these methods is shown to be quite effective as long as one or more known initial match points can be provided. Experiments performed by imaging actual patients under realistic conditions indicate that the 3-point transformation algorithm performs the best overall, achieving an average matching accuracy of 97%. The nature of these algorithms, their relative performance under a range of conditions, and possible methods for improving accuracies are discussed.