Psoriatic plaque segmentation in skin images

Abstract

This paper concerns about automatic psoriatic plaque segmentation in skin images. A novel pixel clustering based image segmentation approach is proposed. The clustering approach works with the circular-linear color spaces which remained unexplored because of the unavailability of appropriate mixture models and hence, the traditional clustering algorithms are incapable of handling such color spaces. This paper makes use of a novel Semi-Wrapped Gaussian Mixture Model (SWGMM) based approach for clustering image pixels. The experiment considers CIE Lch color space which is essentially a circular-linear space. The performance of SWGMM-based segmentation is compared to commonly used Fuzzy C-Means(FCM) and Fuzzy Kernel C-Means (FKCM)-based segmentation methods in linear color space. Results on a newly generated dataset show that SWGMM-based segmentation achieves accuracy of 82.84% in CIE Lch; whereas experiments with three linear color spaces namely, RGB, CIE L*u*v* and CIE L*a*b* show that FKCM and FCM achieve best accuracies of 78.80% and 77.79%, respectively in Lu*v* space. For comparing the clustering-based diseased detection result to a supervised pixel classification based approach, a support vector machines (SVM) based classifier is trained with labelled diseased and normal healthy skin pixels. As the number of images is limited for the present problem, SVM based diseased detection achieves 63.56% accuracy which is significantly lower than the accuracies given by the clustering based approaches.