Automatic Hemangioma Detection Algorithm Using a Cascade of K-Means and Active Contour Model

Although mostly harmless, hemangiomas still need to be monitored and occasionally treated to avoid complications. The method presented for accurately segmenting the hemangioma pixels involves the automatic detection of the number of classes in an initial k-means clustering, followed by binarization, morphological operations and a further adjustment of region of interest using active contours. The method has been tested on a database containing a variety of situations, including multiple hemangioma areas, differently colored and textured skin and intrusive hair. Compared to the results before the addition of active contours, the mean global score shows an improvement of more than 1% (from 96.86% to 97.92%).