Evaluation of color spaces for unsupervised and deep learning skin lesion segmentation

Abstract

The reliability of skin cancer diagnosis is dependent on accurate lesion segmentation. The choice of a color space in most contributions on skin lesion segmentation for melanoma detection are based on qualitative rather than quantitative approaches. User experience and theoretical properties of the color space are the two major factors influencing the choice of the color space. For this reason, it may be difficult to optimize segmentation accuracy. This paper evaluates the discrimination power of 5 color spaces and 16 color channels for two unsupervised approaches and a deep learning approach on the segmentation of skin lesion in dermatoscopy images. 600 dermatoscopy images with different levels of cluttering and occluding objects from two different databases were utilized. This study suggests that no single color space or color channel is most suitable in real-world scenarios. Therefore, this study can be regarded as a general framework to select a single or combination of color channels that will enhance the segmentation accuracy of images with different level of scene complexities and illumination variations.