Automatic segmentation of anterior segment optical coherence tomography (AS OCT) images provides an important tool to aid management of ocular diseases. Having precise details about the topography and thickness of an individual eye enables treatments to be tailored to a specific problem. OCT is an imaging technique that can be used to acquire volumetric data of the anterior segment of the human eye. Fast automatic segmentation of this data, which is not available, means clinically useful information can be obtained without the need for time consuming error-prone manual analysis of the images. This thesis presents newly developed automatic segmentation techniques of OCT images. Segmentation of 2D OCT images is first performed. One of the main challenges segmenting 2D OCT images is the presence of regions of the image that generally have a low signal to noise ratio. This is overcome by the use of shape based terms. A number of different methods, such as level set, graph cut, and graph theory, are developed to do this. The segmentation techniques are validated by comparison to expert manual segmentation and previously published segmentation techniques. The best method, graph theory with shape, was able to achieve segmentation comparable to manual segmentation. Good agreement is found with manual segmentation for the best 2D segmentation method, graph theory with shape, achieving a Dice similarity coefficient of 0.96, which is comparable to inter-observer agreement. It performed significantly better than previously published techniques. The 2D segmentation techniques are then extended to 3D segmentation of OCT images. The challenge here is motion artefact or poor alignment between each 2D images comprising the 3D images. Different segmentation strategies are investigated including direct segmentation by level set or graph cut approaches, and segmentation with registration. In particular the latter requires the introduction of a registration step to align multiple 2D images to produce a 3D representation to overcome the presence of involuntary motion artefacts. This method produces the best performance. In particular, it uses graph theory and dynamic programming to segment the anterior and posterior surfaces in individual 2D images with shape constraint. Genetic algorithms are then used to align 2D images to produce a full 3D representation of the anterior segment based on landmarks or geometric constraints. For the 3D segmentation, a data set of 17 eyes is used for validation. These have each been imaged twice so a repeatability measurement can be made. Good repeatability of results is demonstrated with the 3D alignment method. A mean difference of 1.77 pixels is found between the same surfaces of the repeated scans of the same eye. Overall, a new automation method is developed that can produce maps of the anterior and posterior surfaces of the cornea from a 3D images of the anterior segment of a human eye. This will be a valuable tool that can be used for patient specific biomechanical modelling of the human eye.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:664443 |
Date | January 2015 |
Creators | Williams, Dominic |
Publisher | University of Liverpool |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://livrepository.liverpool.ac.uk/2013626/ |
Page generated in 0.0016 seconds