A non-invasive automated approach for eczema lesions segmentation using colour space normalization

Eczema is a common type of atopic dermatitis. Eczema skin lesions can be identified visually by observing the difference between the colour and texture of the lesions and the normal skin. Dermatologists assess eczema by direct visual assessment and record their observations in specialized forms. These methods are tedious as well as time consuming and introduce inter-rater and intra-rater variability in the results. To make the assessments objective and easier for the dermatologists, we have proposed a framework for the segmentation of eczema skin lesions. Red-Green-Blue (RGB), CIELab and their normalized colour spaces were considered. For segmentation a two step K-means algorithm was proposed. In the 1^st step, a conventional K-means algorithm segments the image into three regions, i.e., skin, lesion, and mixed region. This was followed by a 2^nd K-means segmentation step. The performance of this method was better than the conventional methods. The algorithm was evaluated using 85 eczema images of different severity and grades. To assess the performance of the algorithm, the gold standard segmentation for eczema lesions was manually drawn and verified by a dermatologist. The Green-channel of normalized (CSN-I) RGB colour space provided the best result for a semi-supervised approach giving the segmentation accuracy of 88.28% whereas for fully automated approach a segmentation accuracy of 84.43% was achieved using support vector machine (SVM).