Towards a computer-aided diagnosis system for pigmented skin lesions

Abstract

This paper presents a computer-aided diagnosis system for pigmented skin lesions, with solutions for the lesion boundary detection and for the quantification of the degree of symmetry. Lesion detection results were validated by expert dermatologists, who also provided hand-drawn boundaries of the lesions. These reference boundaries were not used as a gold standard, but were allowed to statistically determine the accuracy of the boundaries provided by computerized techniques. We could show that the dermatologists were not able to reproduce their results, and that the boundaries of any expert taken alone showed higher divergence from those of the set of remaining experts than the automatic techniques we developed. Feature extraction is restricted in this paper to the quantification of degree of symmetry, even though it is clear that many other features will be necessary for a complete diagnosis system. The symmetry quantification step provides a six-dimensional feature vector that can be used to classify pigmented skin lesions as being benign or malignant. We demonstrate that our scheme outperforms methods based on the principal component decomposition, which is widely used for this kind of application.