Overview
Diabetic retinopathy is a leading cause of blindness in working age people in the Western world. Not everyone with diabetes will develop diabetic retinopathy but people with diabetes are invited to have the back of their eyes photographed each year as part of a screening programme to catch it early.
Of the 2.5 million people with diabetes in England, 1.9 million were screened in 2011-12. At the moment, the pictures are checked by people trained to grade them on whether they show diabetic retinopathy and at what stage. This takes time and is expensive.
In this project the student is developing an automatic way to find and classify the condition using a computing technique known as 'deep learning'. They will use around 4,000 images to teach the computer which features turn up in the ones from people with diabetic retinopathy (and aren’t there in people who don’t have the condition). They’ll also teach the computer to grade the images and will finally test it against human graders, to find out how well it does. If successful, the team will look for extra funding to turn the software into a tool that can be used in the clinic.
-
Research update
Researchers have completed the recruitment target of 103 patients. They have completed gradings of the obtained images according to the protocol developed. Researchers have also developed and optimised computer programs for automated assessment of image quality. The preliminary data suggest that images of the optic disc can be obtained with both of the devices being tested, in both dilated and undilated eyes. However, the field of view from undilated eyes is restrictive.
-
Scientific summary
Deep learning for automatic screening of diabetic retinopathy
This project aims to develop a new automated image grading system to provide point of care full disease retinopathy grading results for people with diabetes. Diabetic retinopathy (DR) is a progressive condition that leads to blindness. In the UK patients with diabetes are annually invited to have photos taken of their eyes for the screening of DR severity. Currently classification of DR severity is conducted by graders who manually inspect retinal images. Manual grading is a subjective and costly process. Automated grading is believed to improve screening efficiency and ensure consistency and accuracy of grading. Full disease grading essentially is a challenging imbalanced multi-class classification problem in computer science.
Building on the team's achievements in retinal image mining for eye disease, they are employing deep learning techniques to address this real-world problem further. More specifically, they are: (1) investigating how to learn features to best represent individual images for the accurate grading of DR, (2) building classification models to achieve the performance required for DR screening, (3) using GPU computing technology to speed-up data processing and training process without affecting accuracy, and (4) validating the developed tools using public datasets and their own real screening data so as to identify the best techniques for future clinical evaluation.
On completion of the project, further funding will be sought to support the translation of the developed tools to provide tangible benefits for an enormous number of people who are at risk of sight loss from DR in the UK and beyond.