• Ting Shu, Bob Zhang, and Yuan Yan Tang.
• Department of Computer and Information Science, Avenida da Universidade, University of Macau, Taipa, Macau, China. Electronic address: yb57406@umac.mo.
• Comput. Biol. Med. 2017 Apr 1; 83: 69-83.

IntroductionResearchers have recently discovered that Diabetes Mellitus can be detected through non-invasive computerized method. However, the focus has been on facial block color features. In this paper, we extensively study the effects of texture features extracted from facial specific regions at detecting Diabetes Mellitus using eight texture extractors.Materials And MethodsThe eight methods are from four texture feature families: (1) statistical texture feature family: Image Gray-scale Histogram, Gray-level Co-occurance Matrix, and Local Binary Pattern, (2) structural texture feature family: Voronoi Tessellation, (3) signal processing based texture feature family: Gaussian, Steerable, and Gabor filters, and (4) model based texture feature family: Markov Random Field. In order to determine the most appropriate extractor with optimal parameter(s), various parameter(s) of each extractor are experimented. For each extractor, the same dataset (284 Diabetes Mellitus and 231 Healthy samples), classifiers (k-Nearest Neighbors and Support Vector Machines), and validation method (10-fold cross validation) are used.ResultsAccording to the experiments, the first and third families achieved a better outcome at detecting Diabetes Mellitus than the other two.ConclusionsThe best texture feature extractor for Diabetes Mellitus detection is the Image Gray-scale Histogram with bin number=256, obtaining an accuracy of 99.02%, a sensitivity of 99.64%, and a specificity of 98.26% by using SVM.Copyright © 2017 Elsevier Ltd. All rights reserved.

20 keywords...

hide…