Characterising shape patterns with applications to biomedical image analysis

Author: Amelia Gontar

Gontar, Amelia, 2018 Characterising shape patterns with applications to biomedical image analysis, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact with the details.


The development of shape analysis methods is an important field of study, especially in the context of biomedical image analysis. An understanding of shape patterns provides information on morphology, function, growth, abnormalities, and so on. Various shape analysis methods -- including elliptic Fourier analysis, multiple resolution skeletons, landmark methods and statistical shape models -- have been covered extensively in the literature. However, these methods have been developed to analyse regular shape patterns and do not extend well to objects exhibiting irregular shape patterns, such as those common in biomedical settings.

In this thesis, a method for shape analysis is developed specifically for cases where the shape patterns are highly irregular. The idea of clustered shape primitives is introduced, in which local shape patterns are captured and clustered to form representative shape patterns occurring commonly throughout an object or group of objects. Histograms of the occurrences of the clustered shape primitives are used to characterise the shape patterns, and may be used for classification or for regression to predict a given property of the data.

The method is demonstrated and explored using three example data sets: Saccharomyces cerevisiae yeast colonies exhibiting pseudohyphal growth, cancellous bone in rat tibiae, and marbling in beef. Each of these types of objects exhibit highly irregular shape patterns that cannot be adequately described using existing shape analysis methods, for example because the consistent placement of landmark points between samples is difficult or impossible.

The most significant contribution of clustered shape primitives is that complex shape patterns are learned automatically. There is no need to define any features or landmark points at the outset of the study, which is an improvement on previous methods wherever irregular shape patterns are present. Another advantage of the method proposed here is that the number of features can be kept small, since the most important features are learned and selected automatically. This avoids the need to define a very large number of complex geometric features and ensures that important features are not overlooked -- these are examples of problems that may arise when a large list of features is defined at the outset of the study.

Classification methods based on clustered shape primitives achieved competitive results wherever only binary data was available and the shape patterns exhibited were complex. The method could have an impact on shape analysis in biomedical images, and the current study forms a basis for future work in this direction.

Keywords: shape analysis, dimorphic yeast, pseudohyphal growth, cancellous bone, marbling in beef, clustered shape primitives, shape characterisation

Subject: Mathematics thesis

Thesis type: Doctor of Philosophy
Completed: 2018
School: College of Science and Engineering
Supervisor: Murk Bottema