Computational approach for chronic wound tissue characterization

This paper describes an automated diagnostic system for continuous chronic wound status monitoring. Accurate and periodic wound assessment is important for optimal wound care. Automated wound diagnosis is beneficial for the aging population, to obtain a treatment-related decision for clinicians. Wound healing analysis can be done using image pre-processing, segmentation, and classification, with visual evaluation by a learned clinician. In this paper, we proposed fuzzy c-means clustering for wound image segmentation, in conjunction with the standard computational learning schemes: Linear Discriminant Analysis (LDA), Decision Tree (DT), Naïve Bayesian (NB) and Random Forest (RF). These techniques are useful for classifying the percent of wounded tissue in a segmented region. The color features observed in the imagery were expected to be helpful to enhance resolution. The histogram sampling method was found useful to provide a wider separation between wounded portions in the color space. The overall accuracy of our paradigm was measured with respect to images judged by an expert clinician, who manually traced the wound portions, i.e., for groundtruth images versus segmented wound images. The outcome of the proposed technique provided a 93.75% overall accuracy; whereas, using the Random Forest scheme with Decision Tree, Linear Discriminant Analysis, and Naïve Bayesian, we obtained 84.29%, 85.67%, and 78.66% accuracy, using manual segmentation as groundtruth. The proposed scheme achieved an overall performance comparable to the best results reported in the literature. Keywords: Chronic wounds, Image segmentation, Tissue classification, Feature extraction, Performance analysis