JultikaUniversity of Oulu repository

Abstract

This thesis attempted to optimize a novel GLCM-based 3D Texture Analysis software in terms of its input parameters in order to maximize the early prediction of knee osteoarthritis from 3D DESS MR images. 20 subjects (10 control subjects; 10 progressor subjects) containing image data from baseline and from a 36-month-follow-up were extracted from the Osteoarthritis Initiative database and used as the study dataset. Multiple sets of 3D Texture Analysis were conducted incorporating 22 static and dynamic grey level quantization schemes, 6 bin quantization schemes and 4 offset settings. Cliff’s delta was calculated to measure the effect size between the patient cohorts. Multilayer perceptron, Naïve Bayes and Support Vector Machines were implemented to classify the patients into their respective cohorts and estimate the robustness of the 3D Texture Analysis outputs. The predictions were done using only the baseline data, where all patients showed no signs of osteoarthritis. Maximum achieved robustness was 87%. The 3D Texture Analysis was found to have a high potential for the early prediction of knee osteoarthritis based on the GLCM features and the results outlined the importance of the software’s input parameters.