Classification and Identification of Osteoporosis Cases through X-ray Images

Abstract

Evaluating osteoporosis disease using digital radiography presents a significant challenge in pattern recognition and image processing applications. Textures of the images from the bone micro-architecture of osteoporotic patients and healthy subjects are similar to a great extent, thus enhancing the complexity of the classification of such textures. In this dissertation, we study the effects of the orientation in texture analysis and the performance of LBP, 1DLBP and sLBP. We use both global and local information. Image Projections are used to capture the global information and the local features are obtained using neighborhood operations. We study the effects of using different directions individually using Fourier domain filtering and then quantizing the filtered images. The features extracted are LBP, 1DLBP and sLBP. We study the performance of shift LBP that deals with noise and fuzzy nature of the patterns. Then, we propose a method that combines the 1DLBP (captures global as well as local patterns) and multi resolution gabor filters. Gabor filters have the properties of the spatial locality and orientation discrimination that enables it to capture patterns in multiple orientations. Our proposed schemes are validated using a ten-fold cross validation using KNN and SVM classifiers. The best results obtained with KNN is 69.32% and with SVM is 66.89%. Our results showed that capturing the patterns in all the directions concurrently is much more beneficial than extracting features in a specific single direction. This study helps in providing a better discrimination of two population of OP and CC cases.