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 88% , average specificity 96%). This method learns by example and produces fully automatic algorithms needing no human interaction or parameter tuning, and although complex, runs in approximately 4 seconds.

• Type: Conference
• Year: 2003
• Where: MICCAI
• Authors: Mark E. Roberts, Ela Claridge
• Keywords: Average Sensitivity, Average Specificity, Medical Imaging, MICCAI 2003, Unseen Test Images