ABSTRACT

Adaptive Multi-Scale Texture Analysis

With Application To

Automated Cytology

In this thesis we investigate the application of image analysis and self-adaptive algorithms to the detection of cell abnormalities in cervical smears. Cervical cancer is a preventable disease and, unlike most cancers, can be easily detected by a routine screening test. Current manual screening methods are costly and sometimes result in inaccurate diagnosis due to human error. The introduction of machine-assisted screening will therefore bring significant benefits to the community, by reducing financial costs and increasing screening accuracy.

One of the fundamental weaknesses of research efforts over the last 30 years has been in identifying a robust set of cell descriptors to allow accurate classification of cytological samples. Continuing advances in imaging technology and computing power have provided incremental gains in the diagnostic accuracy of automated cytology systems, however, there is a need for further improvement. The quantitative analysis of cell nuclear texture has shown the most promise in the past, and continues to be a major focus of research efforts around the world. Our motivation for the work in this thesis is a belief that greater effort in texture analysis research will yield further advances of significant benefit.

We investigate the history of texture analysis as applied to automated cytology, and identify Markovian techniques as being powerful methods of analysis. By Markovian methods, we mean those techniques which model the joint or conditional statistical dependence between neighbouring image pixels (Markov chains, Gibbs/Markov random fields, co-occurrence matrices etc.). In this thesis we concentrate on second-order co-occurrence-based techniques---a powerful and computationally light subset of higher-order Markovian approaches. We identify a weakness common to co-occurrence and many other methods of analysis. This weakness is the commonly used technique of applying a set of fixed functions to extract discriminant features. Such functions may provide good general performance across a wide range of texture types, but often fail to capture texture information which is specific to the subset of types being analysed. That is, the methods are globally applicable but are not locally optimised.

We present a theoretical approach to texture classification which is applicable to all texture types but which `adapts' to the specific characteristics of the texture being analysed. The self-adaptive multi-scale techniques based on this approach allow the simultaneous capture of texture characteristics which exist at, and across, several spatial resolutions. We show by a critical appraisal of the presented methods that this approach can provide significant improvements in cell classification accuracy. We also show how the captured characteristics can be used to identify image locations where differences between texture classes occur---something which is generally not possible with other analysis methods. Finally, we demonstrate the broad applicability of our methods by classifying a wide range of texture images from natural, industrial and biological origins.


Ross Walker