Unsupervised segmentation of scaling in 2D psoriasis skin images

Abstract

Psoriasis is a chronic skin disease whose causes are not well understood. There is a lack of acknowledged assessment methods for assessing the efficacy of psoriasis treatment. The search for objective assessment methods has lead to the introduction of numerous computer assisted methods. However, much of the research, has focused on segmenting psoriatic lesions. Scaling, which importantly affects the severity assessment is far less prevalent in the literature.

In this paper, we propose a technique for segmenting scaling from psoriasis skin images. Our technique proceeds by first enhancing the contrast of the scaling with respect to the neighbouring erythema by using a Saliency analysis combined with Gabor feature analysis. Secondly, after removal of erythema in the saliency map, a Markov Random Field (MRF) is used in conjunction with a Support Vector Machine (SVM) to classify the remaining pixels as either scaling or normal skin. The training data for the SVM is gathered automatically from the image being analysed using a soft-constraint K-means clustering. Experimental results indicate that our combination of MRF and SVM performs better than either a SVM or a MRF alone. Moreover, there is no indication that manually selected training sets perform substantially better that the proposed automatic method.