An Efficient Classification for Detecting Diabetes Mellitus and Nonproliferative Diabetic Retinopathy by using PSVM Classifiers

Diabetes mellitus (DM) and its cause’s diabetic retinopathy (DR) are major health problems. To combat this approaching epidemic, proposes a noninvasive automated method to detect DM and NPDR by distinguishing Healthy/DM and NPDR/DM sans NPDR (DM without NPDR) samples using an array of tongue features consisting of color, texture, and geometry. The various tongue features like color, texture, and geometry are extracted from the human tongue can used to diagnose disease. Tongue images were captured using an especially designed in-house device taking into consideration color correction. Each image was segmented in order to locate its foreground pixels. With the relevant pixels located, three groups of features namely color, texture, and geometry were extracted from the tongue foreground. The tongue color gamut represents all possible colors that appear on the tongue surface representing 12 colors. The color features are extracted from the sample tongue image. The texture features are extracted from the tongue image. To extract the texture value of each block, the 2-D Gabor filter is applied. The texture of tongue images separate eight blocks that categorize the nine texture features of the tongue. Here the Healthy samples having a higher texture value and DM samples have lower texture values. Also, 13 geometry features are extracted from tongue images. These features are based on measurements, distances, areas, and their ratios. Finally by applying 34 features, Proximal Support Vector Machine (PSVM) classifier used for classification instead of SVM classifier by comparing with the normal tongue features. Classification was performed between Healthy versus DM in addition to NPDR versus DM-sans NPDR initially using every feature individually (from the three groups), followed by an optimal combination of all the features. A dataset consist of 296 Diabetes Mellitus samples consisting of 267 DM-sans NPDR and 29 NPDR processed among three features and identify the accuracy of 95% result respectively.