###Using physics-informed machine learning to understand the progression of diabetic retinopathy
Project ID: 2228bd1052 (You will need this ID for your application)
Research Theme: Healthcare Technologies
UCL Lead department: Division of Medicine
Lead Supervisor: Simon Walker-Samuel
Project Summary:
Increasing diabetes incidence and an aging population have placed an unprecedented burden on clinical ophthalmology. With an anticipated 60% increase in demand for services over the next twenty years, this will necessitate new ways of working. Equally, a range of new and improved retinal imaging technologies have been developed that have the potential to improve both patient monitoring and diagnosis, but which require in-depth expert interpretation. To address these challenges, this project will combine mechanistic, biophysical modelling of the retina with machine learning tools to assist ophthalmologists with the interpretation of image data from wide-field, colour retinal photography and optical coherence tomography angiography (OCT-A). The platform will also be used for treatment planning and optimisation.
The project will draw on the large database of images from these new modalities uniquely available at Moorfields Eye Hospital. A key novelty of the machine learning approach used here will be a greatly reduced reliance on access to curated, labelled image data. Instead, we will use a combination of mechanistic computational modelling of the retina, alongside machine learning with generative adversarial networks (GANs), to automatically segment images, identify pathology and predict retinal oxygenation.
This will draw on our mathematical model of human retinal vascular structure and fluid dynamics (blood flow, oxygenation), based on our REANIMATE framework (d’Esposito et al, Nature Biomed Eng, 2018). Unpaired cycle-consistent adversarial networks such as the cycleGAN algorithm (Zhu et al, arXiv, 2018) enable mappings between domains to be performed without the need for ground-truth labels. We will use this novel approach to identify features in retinal photographs and OCT-A images, by creating a library of synthetic retinal structures.