Overview
Retinopathy of prematurity is a condition that can affect babies who are born early, before blood vessels in the eye have finished growing. The light-sensitive layer of the eye (the retina) needs oxygen and nutrients from the blood. So if the blood vessels that supply aren’t formed before birth, new vessels grow once the baby is born. But these are fragile and unhealthy and can cause damage that leads to permanent blindness in the most severe cases.
If the condition is caught early, laser treatment is very successful at saving the sight of most affected babies. Premature babies should be screened regularly until the danger of retinopathy of prematurity has passed. This should be done by trained professionals with the right equipment, but the condition is on the rise as modern medicine means that more premature babies are surviving. There are not always enough qualified people available to examine all the babies at risk, especially in middle income countries.
In this project the research team is developing a computer system that can monitor premature babies for signs of retinopathy of prematurity. The computer programme is learning what images of the retina look like at different stages of development. The images are from babies who either do or do not go on to develop the condition.
Eventually the research team would like the programme to be able to predict what a retina might look like in future in the same sort of way that a computer that learnt about ageing might be able to take a picture of a young face and show you what it might look like in the future. And by the end of the project the team hopes to have a system that can help health professionals monitor the condition with more confidence so that treatment is given at the right time when needed to save a baby’s sight.
-
Scientific summary
Computational algorithms for ROP screening and prevention.
This project will design accurate, informative computational algorithms for retinopathy of prematurity (ROP) screening. On the foundation of a unique retinal vessel detection system developed previously by the team, this project will build the first generative probabilistic model for ROP screening, resulting in a temporal statistical model of how ROP progresses, whilst modelling retinal development in preterm infants without ROP.
Promising pilot studies from the team show a 70%+ recognition rate for simple binary tasks using supervised learning models to convert previously verified clinically meaningful vessel-analysis results into disease grade predictions. This figure will be improved on with attribute-based measurements and a temporal model. A photo quality regressor using a support vector machine will be developed to understand good and bad quality images appropriately. An initial generative model will be developed using a supervised learning model – developing continuous scale-grade predictions. Attribute enhanced models will search for features within the images invariant to pose and a generative temporal model will result. Outcomes of this newly developed tool for one-shot assessment of ROP will be rigorously evaluated throughout the process, comparing output to expert clinician gradings of the same sets of images.
Implementation of the resulting software algorithms into cotside retinal imaging cameras already in use in clinical practice would provide reassurance for clinicians and could potentially be used as a stand-alone screening tool in areas lacking on-site ophthalmic expertise where there is an epidemic of childhood blindness from this treatable condition.