Spectral Clustering for Unsupervised Segmentation of Lower Extremity Wound Beds Using Optical Images.

Chronic lower extremity wound is a complicated disease condition of localized injury to skin and its tissues which have plagued many elders worldwide. The ulcer assessment and management is expensive and is burden on health establishment. Currently accurate wound evaluation remains a tedious task as it rely on visual inspection. This paper propose a new method for wound-area detection, using images digitally captured by a hand-held, optical camera. The strategy proposed involves spectral approach for clustering, based on the affinity matrix. The spectral clustering (SC) involves construction of similarity matrix of Laplacian based on Ng-Jorden-Weiss algorithm. Starting with a quadratic method, wound photographs were pre-processed for color homogenization. The first-order statistics filter was then applied to extract spurious regions. The filter was selected based on the performance, evaluated on four quality metrics. Then, the spectral method was used on the filtered images for effective segmentation. The segmented regions were post-processed using morphological operators. The performance of spectral segmentation was confirmed by ground-truth pictures labeled by dermatologists. The SC results were additionally compared with the results of k-means and Fuzzy C-Means (FCM) clustering algorithms. The SC approach on a set of 105 images, effectively delineated targeted wound beds yielding a segmentation accuracy of 86.73 %, positive predictive values of 91.80 %, and a sensitivity of 89.54 %. This approach shows the robustness of tool for ulcer perimeter measurement and healing progression. The article elucidates its potential to be incorporated in patient facing medical systems targeting a rapid clinical assistance.