Texture Analysis of Bone Mineralisation Surfaces

Reid, Carol Anne (1995) Texture Analysis of Bone Mineralisation Surfaces. PhD thesis, University of Glasgow.

Abstract

A better understanding of the process of bone mineralisation is needed before effective treatments for bone diseases such as osteoporosis can be found. Bone growth is essentially a surface process, and because of this, scanning electron microscopy has proved of great value in imaging large areas of growth surface. Previous studies have identified six mineral surfaces, each associated with a known metabolic growth state, and each characterised by a particular surface texture. It is desirable to have quick and accurate methods for recognising these surface types and evaluating them quantitatively. The objective of this study was to develop algorithms to segment and classify, automatically, the surface textures found in scanning electron microscopy images of rats' bones. This thesis describes the techniques used, their success and their limitations. Chapter 1 is a brief introduction to the physiology of bone and the use of the scanning electron microscope to study bone surfaces. It explains the background to the study and describes the texture types identified in bone surfaces. It also describes the preparation of specimens for the scanning electron microscope, how the images are captured by the SEM and the transfer of the images to disk for use on a Sun SPARCstation. Chapter 2 introduces the idea of texture, and reviews methods previously used to classify texture. The techniques are divided into 3 groups - statistical, structural and modelling. The statistical techniques reviewed include variations based on co-occurrence matrices, grey-level difference statistics and run-length matrices. Structural methods mentioned include a linguistic approach to extracting primitives, an approach based on grey-level thresholding and an edge-based approach. Modelling techniques include random mosaic models, Markov and Gibbs random field models, and random walk models. Each method is briefly described and results, where relevant, are quoted and discussed. In chapter 3 the classification techniques described in chapter 2 are tested on the scanning electron microscopy images of rats' bones. To accelerate the investigation 64x64 windows are extracted from the larger 512x512 images. Image standardisation techniques are also discussed and investigated. None of the statistical classification methods works as well as stated in published results, where, on the whole, images from Brodatz' (1966) texture album were used. The best classification technique for the images in this study is a maximum likelihood classifier based on the co-occurrence matrix without feature extraction. High rates of classification (88%-100%) are obtained on training sets, although results are poorer when applied to test sets. Chapter 4 deals with the problem of segmenting the images. The idea of edge- detection is introduced and texture segmentation techniques are described and investigated. Initially the results in this study are worse than those in published literature but by combining and modifying some of the techniques the segmentation is greatly improved. The chapter also examines automatically thresholding the images to highlight edges. Results of the fully automatic segmentation process are shown. Chapter 5 brings together the work of chapters 3 and 4. It deals with the larger 256x256 images, the drawing of boundaries between texture types using the methods of chapter 4 and subsequent classification of each region into one of 5 types. This chapter also discusses the detection of vascular channels and bone cells and illustrates some of the difficulties in segmenting and classifying the images. Examples are given which show that, to a certain extent, the objective has been achieved. Boundaries can be found between formative and resorptive texture types and around vascular channels. Windows can then be classified, with reasonable accuracy, into one of the 5 texture types. Chapter 6 discusses the limitations of the methods used and the problems of image capture. It also discusses possible ways of improving the results.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Additional Information: Adviser: Hugh Elder
• Keywords: Computer science, Biostatistics
• Date of Award: 1995
• Depositing User: Enlighten Team
• Unique ID: glathesis:1995-76444
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 19 Nov 2019 14:20
• Last Modified: 19 Nov 2019 14:20
• URI: https://theses.gla.ac.uk/id/eprint/76444