Use this URL to cite or link to this record in EThOS:
Title: Deep learning for diabetic retinopathy diagnosis & analysis
Author: Pratt, H.
ISNI:       0000 0004 7964 2478
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
The diagnosis of diabetic retinopathy (DR) during large-scale diabetic screening is important to prevent sight loss in a significant proportion of the working population. The early detection of disease and quantification of disease progression is vital in order to prevent future loss of vision. Diagnosis of DR is performed through medical image analysis. After the success of deep learning in other real-world applications, deep learning is also providing solutions with good accuracy for medical image analysis and is seen as a key method for future applications in the health sector. Current DR image analysis methods offer some automation of the feature extraction process for features of DR but do not utilise the benefits of deep learning. The application of initial deep learning methods to the diagnosis of DR presented in this thesis show promising initial results on referable DR diagnosis. The extension of these initial methods to more complex deep learning models and correlating multiple eye information shows that deep learning can obtain a state-of-the-art classification for the referral of DR. The classification of DR into more granular diagnosis also achieves reasonable accuracy. This thesis also presents a method of training on large images using the Fourier domain in order to process medical images more efficiently during deep learning training. The speed of diagnosis that deep learning provides suggests it is viable for point of care diagnosis. The visualisations techniques presented in this thesis, for both referable and multi-class diagnosis, offer clinical insight into the models for predictions. The model visualisations presented give the clinicians, or graders, the information in order to increase their diagnostic speed through highlighting regions of disease and suggesting the severity level of disease. Hence, the work presented shows the benefits of implementing deep learning analysis to the problem of DR diagnosis either for timely referrals or for grading support.
Supervisor: Zheng, Yalin ; Coenen, Frans ; Harding, simon Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral