Volumetric texture analysis based on three dimensional Gaussian Markov random fields

Abstract

The analysis of rich information provided by volumetric data is paramount for developing ro- bust applications, especially in the medical domain where volumetric images are common. The volumetric texture is a fundamental characteristic that plays a vital role in understanding and analysing volumetric images. The current approaches to volumetric texture analysis commonly extract texture features from two-dimensional (2D) slices of the volumetric images while ignoring the valuable information provided by the third dimension in volumetric data. This approach is therefore not capable of efficiently analysing textures in emerging three-dimensional (3D) images, which results in information loss. Consequently, the development of new methods to analyse volu- metric texture is essential because of recent advances in 3D technologies of imaging, especially medical imaging. Among texture analysis methods, model-based Gaussian Markov random fields (GMRFs) are emerging as a choice for modelling texture and have been used to characterise textures in 2D images. Extending the use of GMRF to characterise textures in volumetric images has not received much attention, and therefore exploring this would be beneficial in many applications, especially in medical image analysis applications in which rich texture information is available in the form of volumetric images. This thesis proposes a new method based on GMRF to characterise textures in volumetric images. The 3D GMRF is developed to extract texture features from 3D patches of volumetric images taking into account the information found in the third dimension. The features extracted by the proposed method are then employed for volumetric texture classification, segmentation and texture-based region tracking. The challenges that arise while dealing with textures in 3D images, such as the growth of sampling points, achieving rotation invariance, and the high dimensionality of feature vectors, are investigated and appropriate solutions are proposed. The proposed methods demonstrate a higher performance compared to other methods through a comparison evaluation carried out on a synthetic textures database. Moreover, the methods pro- posed here are exploited to solve three real-world problems by attempting to diagnose emerging serious chronic obstructive pulmonary disease (COPD), classifying subjects with lung cancer, and analysing cilia motion using clinical datasets. The proposed methods do not require extensive training data or powerful hardware, which makes them suitable for medical image applications where datasets are usually small.