Image Analysis for Segmentation of Psoriasis Lesion

Psoriasis, a hereditary inflammatory skin condition that currently affects 2 - 3 % of the
world's population, is marked by reddish, scaly rashes or lesions covered with scabs of
dead skin. lbis dermatosis is currently not curable but the symptoms can be effectively
controlled through an accurate assessment scheme and well-integrated medical care
therapy. The National Psoriasis Foundation Medical Board has published a guideline
that categorizes the severity of psoriasis - mild, moderate and severe, each
characterized by the percentage of lesions on an individual's body surface area.
However the caveat remains that the distinction between the different categories of
severity is largely influenced by the clinical practitioner's subjectivity. As a result,
PASI scoring is introduced. PASI (Psoriasis Area and Severity Index) is currently the
gold standard method to measure psoriasis severity by evaluating the area, erythema,
scaliness and thickness of the lesions. These 4 parameters require the lesions to be first
segmented from the skin patches before they can be assessed and scored individually.
lbis report thus investigates digital image analysis techniques to segment psoriasis
lesions. In this work, 90 patients are categorized into groups of differing skin tones
based on mean values in the L * component. The 1000 new colours obtained through
clustering of pixel values in the R, G and B component are used to construct three
different skin lesion models. The validation of the three models is done by comparing
the mean values of constructed models and original image, which are found to be the
same. For segmentation of skin into involved and non-involved regions, iterative
thresholding and Otsu's method are applied in 3 colour spaces, namely, I 1hh. CIE L *
a* b* and HSI. The average segmentation errors in the 3 colour spaces are then
compared to select the best colour channel in which to perform the segmentation for
either thresholding. The specificity and sensitivity analysis with the accompanying
Type I error and Type II error are conducted as well. From the segmentation results of
skin lesion models, it is found that segmentation in the h colour channel (for fair skin
tone), !3 and b colour channels (for middle skin tone) and lz, b and S colour channels
(for dark skin tone) yields high accuracy. The same thresholding method in
corresponding colour channels is then applied on 20 real skin samples. The segmented
images are compared with the reference images to measure the accuracy of the proposed
lesion segmentation method in different colour channels. Out of 20 cases, the
segmentation method achieved accuracies of higher than 95% for 19 cases. The lowest
accuracy obtained is for a particular skin-lesion patch with accuracies of 92 - 93%. The
lower accuracy is due to wrinkled skin areas which have been exposed to unequally distributed
light leading to misclassification as lesions. For each different skin tone, the
overall accuracy results show that the proposed colour channels are appropriate and
accurate to carry out the Otsu' s method for segmentation of psoriasis lesions.