Integrating Deep Learning and Image Processing Techniques into a Hybrid Model for Glaucoma Detection
Date
2021-06-23
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Glaucoma is the world's second-leading condition of irreversible vision loss after cataracts, accounting for 12% of annual cases of blindness. Glaucoma is a group of diseases that causes the degeneration of the retinal ganglion cells (RGCs). The death of RGCs leads to structural changes to the optic nerve head and the nerve fiber layer which leads in turn simultaneous functional failure of the visual field. These two effects of glaucoma cause peripheral vision loss, and, if left untreated, eventually blindness. Apart from early detection and treatment, no cure for glaucoma exists. Early detection is dependent on manual observation of patient's clinical data, including retinal images, OCT, and visual field test, by an ophthalmologist, which is costly and may be prone to error. As a result, most patients remain undiagnosed or improperly diagnosed, such that glaucoma progresses leading to more irreversible vision loss before it is detected. A need to enhance the diagnosis of glaucoma and thereby help to decrease blindness thus exists. This diagnosis can be effectively aided by investigating retinal images (also called fundus images) of the interior of the eye. Advances in the fields of deep learning and digital imaging have increased the potential for extracting information from the fundus images for glaucoma assessment. In this thesis, my work focuses on approaches that may help optometrists/ophthalmologists when assessing the health of an eye based on the fundus images. This lead to improving the detection rate of glaucoma and this would help reducing vision loss from progressing by early treatments. Information extracted from retinal images can be of great help when diagnosing glaucoma as noted above. One of the main informative features of the eye is the optic disc. Images of this disc may be isolated from fundus images using computational tools and hence it can be monitored and evaluated for progression when glaucoma is suspected. Other features such as changes in the optic disc region can also be used as one of the indicators when diagnosing glaucoma. For instance, the cup-to-disc ratio can be used to detect the level of intra-ocular pressure in the eye. Other indicators such as the vessels in the eye can be monitored in a similar manner. In this thesis, I develop a hybrid approach that uses various retinal structures for glaucoma detection. A deep learning-based approach for segmenting the fundus vessels, disc, and cup from retinal images is proposed. I take into consideration the issue of the limited number of retinal images available along with the variability of such images as obtained from various sources. Various features, such as the cup-to-disc ratio, are utilized to classify whether a retinal image is glaucomatous or not.
The main contributions in this thesis can be enumerated as follows:(1) publishing new datasets that can be of great help for researchers working in this field; (2) development of a robust segmentation approaches that can also be helpful when working with other retinal conditions such as diabetic retinopathy; (3) development of a hybrid feature extraction approach from the segmented objects which was utilized by the classifier for glaucoma detection; (4) development of a decision support-based approach that can be the basis of a platform that ophthalmologist/optometrists can utilize when diagnosing glaucoma; and (5) The developed approach can be utilized in the field of telemedicine especially in developing countries where resources are very limited. Moreover, with the advancement in the field of the portable retinal cameras, it is possible to integrate the proposed approach with these devices to facilitate the diagnosis of glaucoma and improve the detection rate, especially in developing countries.
Description
Keywords
Segmentation, Deep Learning, Retinal Disc, Glaucoma, Retinal Vessels, Machine Learning, Medical Image Analysis, Segmentation, Retinal Cup, Dataset, Diagnosis, Domain Shift, Classification, Convolutional Neural Network, Transfer Learning, Data Augmentation
Citation
Sarhan, A. (2021). Integrating Deep Learning and Image Processing Techniques into a Hybrid Model for Glaucoma Detection (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.