Automated Segmentation of Retinal Optical Coherence Tomography Images
MetadataShow full item record
Aim. Optical Coherence Tomography (OCT) is a fast and non-invasive medical imaging technique which helps in the investigation of each individual retinal layer structure. For early detection of retinal diseases and the study of their progression, segmentation of the OCT images into the distinct layers of the retina plays a crucial role. However, segmentation done by the clinicians manually is extremely tedious, time-consuming and variable with respect to the expertise level. Hence, there is an utmost necessity to develop an automated segmentation algorithm for retinal OCT images which is fast, accurate, and eases clinical decision making. Methods. Graph-theoretical methods have been implemented to develop an automated segmentation algorithm for spectral domain OCT (SD-OCT) images of the retina. As a pre-processing step, the best method for denoising the SD-OCT images prior to graph-based segmentation was determined by comparison between simple Gaussian filtering and an advanced wavelet-based denoising technique. A shortest-path based graph search technique was implemented to accurately delineate intra-retinal layer boundaries within the SD-OCT images. The results from the automated algorithm were also validated by comparison with manual segmentation done by an expert clinician using a specially designed graphical user interface (GUI). Results. The algorithm delineated seven intra-retinal boundaries thereby segmenting six layers of the retina along with computing their thicknesses. The thickness results from the automated algorithm when compared to normative layer thickness values from a published study showed no significant differences (p > 0.05) for all layers except layer 4 (p = 0.04). Furthermore, when a comparative analysis was done between the results from the automated segmentation algorithm and that from manual segmentation by an expert, the accuracy of the algorithm varied between 74.58% (layer 2) to 98.90% (layer 5). Additionally, the comparison of two different denoising techniques revealed that there was no significant impact of an advanced wavelet-based denoising technique over the use of simple Gaussian filtering on the accuracy of boundary detection by the graph-based algorithm. Conclusion. An automated graph-based algorithm was developed and implemented in this thesis for the segmentation of seven intra-retinal boundaries and six layers in SD-OCT images which is as good as manual segmentation by an expert clinician. This thesis also concludes on the note that simple Gaussian filters are sufficient to denoise the images in graph-based segmentation techniques and does not require an advanced denoising technique. This makes the complexity of implementation far more simple and efficient in terms of time and memory requirements.
Cite this version of the work
Priyanka Roy (2018). Automated Segmentation of Retinal Optical Coherence Tomography Images. UWSpace. http://hdl.handle.net/10012/13623