Deep Learning for the automated prediction of diabetic retinopathy progression
This PhD project will create longitudinal Deep Learning (DL) models capable of predicting the risk of Diabetic Retinopathy (DR) progression over a 1 to 3-year period. Earlier identification may prompt proactive lifestyle or medical interventions to mitigate the risk of developing vision-threatening DR. DL models may uncover new high-risk features of DR progression within retinal images and facilitate individualised screening intervals that target resources more appropriately.
This proposal is timely and with great potential impact, as DR is estimated to affect a third of the 422 million people living with diabetes worldwide, of whom 3.8 million have vision-threatening DR. The number of diabetic people in the UK is reported to be ~4.7 million (6.6% prevalence) and is predicted to rise to >5 million by 2025.
This PhD project will leverage the UK’s DR screening programme, the largest in the world, with >2 million patient episodes per year over 10 years, each with retinal imaging. Starting from state-of-the-art results on DL for prediction of referable DR and maculopathy (approximately 95% sensitivity/specificity), this translational PhD project will investigate DL architectures tailored to clinical data that represent unbalanced classes. The PhD student will investigate hard-example mining approaches, combination of multiple datasets for network learning bootstrapping, and application specific data augmentation. In collaboration with the London AI Centre for Medical Image Analysis and Value-Based Healthcare, hosted at King’s, the project has the potential to develop ground-breaking clinically relevant AI to determine DR progression and save sight.
Disease progression, multi-modal data