OCTage: monitoring the ageing brain via Optical Coherence Tomography of the eyes

Lead Research Organisation: Newcastle University
Department Name: Biosciences Institute

Abstract

Age-related changes in the brain are one of the most important determinants of how our lives change with age. The only part of the brain that can be viewed directly and non-invasively is the retina at the back of the eye. The retina is actually an outgrowth of brain tissue, and thus changes in the brain due to age and disease are often reflected in the eyes.

Optical Coherence Tomography (OCT) is a relatively new technique which can image the 3D structure of living retinal tissue in near microscopic detail. OCT is already widely used to diagnose, assess and monitor eye conditions such as glaucoma, as well as monitor how systemic conditions such as diabetes are affecting the eyes. However, we believe that OCT imaging of the eyes has much greater potential to monitor brain health as well. We already know that retinal structure reflects demographic factors such as age, sex and ethnicity, and that particular neurodegenerative conditions, such as Alzheimer's and Parkinson's disease, are associated with changes in the thickness of the different tissue layers within the retina. We are excited by the possibility of using OCT scans to monitor brain health as people age.

Achieving this will require detecting and interpreting very subtle changes in retinal structure, far harder to detect than the gross changes caused by eye diseases like glaucoma or age-related macular degeneration. Changes of concern will have to be discriminated from normal variability within diverse populations. And there is already a shortage of ophthalmologists trained to make even relatively straightforward diagnoses from OCT images.

We propose to address this by using artificial intelligence (AI) techniques to recognise and characterise the changes that occur in the retina as people age, both in healthy ageing and in the presence of pathologies such as Parkinson's. Our vision is that one day, people could be offered an OCT scan as part of over-50s health check. These scans are cheap (around £30), quick (under two minutes), contact-free and completely painless. The scan could then be run through automated software which, as well assessing eye health, would assess whether the person was on course for a healthy old age, whether they might benefit from early interventions (e.g. better diet or exercise) or whether they should be referred for investigations for conditions such as Parkinson's. This screening could be delivered by high-street optometrists.

Achieving this vision will require overcoming logistical, ethical, technical, and scientific challenges. For the AI to learn how to interpret retinal images correctly, it will need training on very large amounts of diverse, high-quality, clinically-labelled data in order to ensure robustness and reliability across diverse populations.

The project therefore has 3 parts. First, we will expand our existing dataset of OCT scans, drawing on archives from local NHS hospitals and existing national scientific infrastructure such as the UK Biobank or Health Data Research UK. We have also partnered with Specsavers to collect scans in their optometry stores.

Second, we will design AI methods to learn from this dataset, using techniques geared to the different quality and quantity of data available. For example, large datasets can teach the AI about variability within the population, even with little clinical information; conversely, smaller datasets with detailed clinical information can then suggest how to interpret this variability. We will produce two key pieces of software. One, OCTageNet, will estimate a person's age from their eyes. Where this is older than their actual age, it may suggest that they are not ageing healthily and that intervention could be helpful. The other, OCTagePath, will aim to detect early signs of particular conditions such as Parkinson's.

Third, we will engage with public, patients and stakeholders to better understand concerns and barriers around screening for brain health

Technical Summary

Our first task will be to assemble a large dataset for training. In Workstream 1, we will exploit 3 sources: (1) OCT images from archival clinical data in our own local Trusts, annotated with high-quality clinical information from medical records; (2) much larger datasets from sources such as UK Health Data Research, with sparser clinical information, and (3) OCT scans of BioResource volunteers, performed specifically for the project by our partners Specsavers.

For our AI work in Workstream 2, we will adopt a two-stage approach aimed at making best use of all sources of data, those with sparse clinical information as well as those with detailed records. First, we will train a network, OCTageNet, to predict a person's age given their sex, OCT scan of the macula, a fundus image, and ethnicity where available. For this we will be able to use supervised learning on a very large dataset (100,000s of eyes). This task will encourage OCTageNet to learn a suitable "landscape" within which to represent the features of the ageing retina, including those with the full range of pathologies represented in our dataset. We will also employ a custom loss function to promote the generation of realistic trajectories through this space, capturing how retinal structure changes with age and disease course, using the subset of patients for whom longitudinal data exist.

We will then use transfer learning on OCTageNET, along with curated datasets of individuals with diagnoses such as Parkinson's, multiple sclerosis, diabetes, glaucoma etc, to train second-stage networks, OCTagePath, to recognise particular pathologies affecting the retina. We will examine the accuracy of OCTagePath on individual diagnoses, and also how accuracy is degraded by multi-morbidity.

In parallel, in Workstream 3, we will continue engaging with patients, public and other stakeholders in order to understand their needs, concerns and any barriers to eventual roll-out of screening technology.

Publications

10 25 50
 
Title Retinal age prediction model 
Description A computer algorithm based on deep learning technology to predict the age of a person based on a retinal image captured by a OCT scan. The refinement of this model is ongoing. Based on evaluation on middle-aged people from the UK Biobank the current level of performance of this model is ~3MAE, i.e. an average error in making predictions of +- 3 years. 
Type Of Material Computer model/algorithm 
Year Produced 2024 
Provided To Others? No  
Impact The effort of improving this model is ongoing, and we have not explored yet its potential impact.