An artificially evolved vision system for segmenting skin lesion images

Abstract

We present a novel technique where a medical image segmentation system is evolved using genetic programming. The evolved system was trained on just 8 images outlined by a clinical expert and generalised well, achieving high performance rates on over 90 unseen test images (average sensitivity 97%, average specificity 81%). This method learns by example and produces fully automatic algorithms needing no human interaction or parameter tuning, which although complex, runs in approximately 4 seconds. © Springer-Verlag Berlin Heidelberg 2003.