Disease Detection Using Tongue Geometry Features with Sparse Representation Classifier

Abstract

In this paper we propose a method to distinguish Healthy and Disease individuals through tongue image analysis, specifically via tongue geometry features with Sparse Representation Classifier (SRC). After a tongue is captured using our non-invasive device, it is first segmented to remove its background pixels. Thirteen geometry features based on areas, measurements, distances, and their ratios are then extracted from the tongue foreground pixels. These features then form two sub-dictionaries in the SRC process, a Healthy geometry feature sub-dictionary, and Disease geometry feature sub-dictionary. Experimental results are conducted on a dataset consisting of 130 Healthy and 130 Disease samples. Using all thirteen geometry features SRC achieved a sensitivity of 86.15%, a specificity of 72.31%, and an average accuracy of 79.23% at Healthy vs. Disease classification.