Fraz, Muhammad Moazam (2013) Retinal image segmentation and quantification of vessel width in non-standard retinal datasets. (PhD thesis), Kingston University, .
Abstract
The human retina has the potential to reveal important information about retinal, ophthalmic, and even systemic diseases such as diabetes, hypertension, and arteriosclerosis. Automatic quantification of retinal vessel morphology and width is considered as a first step in computer assisted medical applications related to diagnosis and treatment planning. This work aims to quantify the blood vessels in noisy and pathological retinal images of school children with uneven illumination and containing complex vessel profiles. In this thesis, we have presented two methodologies of retinal vessel segmentation and an algorithm for vessel width measurement. The unsupervised method of retinal segmentation is based on detection of vessel centrelines and followed by computing the vessel shape and the orientation map using morphological bitplane slicing. A supervised method for segmentation of blood vessels by using an ensemble classifier of boosted and bagged decision trees is also presented. The feature vector encodes information to successfully handle both normal and pathological retinas with bright and dark lesions simultaneously. The obtained performance metrics illustrate that this method outperforms most of the state-of-the-art methodologies of retinal vessel segmentation. The method is computationally fast in training and classification and needs fewer samples for training than other supervised methods. It is training set robust as it offers a better performance even when it is trained and tested on different sets of retinal images. A new public database of the retinal images taken from multi-ethnic school children is presented along with the ground truths of vessel segmentation and width measurement. We have also introduced a robust and accurate methodology for measuring the calibre of vessel segments in retinal images of multi-ethnic children. The vessel centrelines are detected from the vessel probability map image resulting from ensemble classification. The vessel branch points and crossovers are identified and removed from the vessel centreline image to obtain vessel segments followed by computing the local vessel orientation of the vessel segments. The width of each vessel segment is estimated using a two dimensional model with incorporated Gaussian (for ordinary vessels) as well as Difference of Gaussian profiles (for vessels with a central reflex). The automated methods for quantification of retinal vessel morphology and width may be used as an alternative to the time consuming subjective clinical evaluation for monitoring the progression of retinopathies and their association with normal and abnormal vascular patterns. This may enable a quick diagnosis, treatment availability, prognosis, and facilitation of clinical heath-care procedures in remote areas.
Actions (Repository Editors)
 |
Item Control Page |