Digital imaging, AI advance investigation of systemic disease through the retina
Click Here to Manage Email Alerts
The emergence of high-resolution imaging modalities and the progress of artificial intelligence, in particular deep learning, have opened up new pathways to identify signs and predictors of systemic disease through the retina.
“The concept that the eye is a window to systemic health has been proposed since the 1850s, when the invention of the ophthalmoscope made it possible to observe correlations between damage in the retina and kidney disease, high blood pressure and heart problems,” Tien Yin Wong, MD, PhD, said.
Source: Tien Yin Wong, MD, PhD
A second phase began in the late 1990s and early 2000s with the advent of digital fundus photography, marking the transition from subjective qualitative assessment to objective quantitative assessment of the retina.
Since about 2015, with the application of AI and deep learning methods to retinal image analysis, a third phase has begun, “the most exciting, because within milliseconds we can analyze the retina and, in theory, predict the risk for specific systemic diseases from digital retinal images,” Wong said.
Pearse A. Keane, MD, FRCOphth, and his group at Moorfields Eye Hospital and University College London, coined the term “oculomics” for the application of deep learning to retinal imaging.
“For more than a century, we have known that signs of systemic disease are visible in the retina. But what is different now is that we can combine advanced retinal imaging, big data and AI to identify biomarkers of systemic disease. The AI revolution offers us opportunities to enhance our understanding of eye-body relationships and to develop new predictive approaches for complex age-related disorders,” he said.
A deep learning algorithm for CVD
Cardiovascular disease (CVD) is a main cause of death globally. Identifying people at risk for cardiovascular events would help prevention through lifestyle-related interventions and proactive administration of risk-reducing therapies.
“The retinal vessels are the only small blood vessels we can see in the body without needing a knife or a scalpel,” Wong said.
Retinal fundus photography and OCT angiography (OCTA) are noninvasive, accessible technologies that could potentially be used for screening patients at high risk for future clinical cardiovascular and cerebrovascular events.
“The current imaging modalities of CVD, such as cardiac CT scan and MR angiography of the brain, are expensive, more invasive and not easily available. In one of our studies, we showed that retinal photographs provide as much information as a cardiac CT calcium score test, which is only available in major hospitals. We could replace very expensive tertiary level modalities with a cheaper, more accessible retinal imaging modality, potentially available in primary care settings,” he said.
With his research group at Singapore National Eye Centre, Wong developed and validated a deep learning-based coronary artery calcium (CAC) score from retinal photographs (RetiCAC) and used this for cardiovascular risk stratification.
A total of 216,152 retinal photographs from five independent data sets from South Korea, Singapore and the U.K. were used to train and validate the algorithms.
“RetiCAC outperformed all single clinical parameter models in predicting the presence of CAC. It showed comparable performance to current CT scan in predicting cardiovascular events and incremental prognostic performance over a commonly used cardiovascular risk prediction model, the pooled cohort equation model, in borderline-risk and intermediate-risk cases,” Wong said.
However, because in most developed countries the elderly populations are not as sick as they were in the past, subtle changes in the retinal vessels are now not easily identified by direct observation of individual ophthalmologists.
“Nevertheless, AI can still pick up the subtle differences between normal and abnormal retina, the retina of a healthy person vs. a person at risk of cardiovascular disease,” Wong said.
In a landmark study by Poplin and colleagues, deep learning models trained on data from 284,335 patients and validated on two independent data sets totaling 13,025 patients were able to predict cardiovascular risk factors such as age, gender, smoking status, systolic blood pressure and major cardiovascular events.
The AlzEye project
Poplin’s study inspired the AlzEye Study, a large-scale study in the U.K. in which patterns of retinal change were associated with the development of dementia.
“We got the idea to look at all forms of retinal imaging captured over the last 10 years at Moorfields Eye Hospital, more than 2 million retinal photos and scans of over 250,000 people, and to link these data to the NHS database on people who developed CVD, Alzheimer’s disease and other forms of dementia. We found that about 10,000 of those patients went on to have a stroke, about 12,000 went on to have a heart attack, and about 13,000 went on to develop dementia,” Keane said.
The focus of the project, led by Keane and Siegfried Wagner, was the development of a screening tool for early detection of dementia, particularly Alzheimer’s disease, based on retinal biomarkers.
The Rotterdam Study published in JAMA Neurology in 2018 had previously shown an association between retinal nerve fiber layer thinning and the development of an increased risk for dementia, including Alzheimer’s disease. Out of 3,289 individuals who received OCT scanning, 41 already had dementia and 86 developed dementia during the follow-up. However, due to the relatively small number of incident cases, the study was not powered to perform prediction modeling for dementia.
“Our large data set will allow us to use AI deep learning to identify individuals at risk of developing dementia. This is something that’s going to be really exciting in the next 12 to 18 months,” Keane said.
The Duke iMIND research team
The retina originates from the diencephalon during embryonic development. It is therefore considered as an extension of the central nervous system and an extension of the changes that would be seen in the brain in any sort of neurodegenerative condition that affects the brain, Dilraj Grewal, MD, said.
There is also a vascular component to the retina that mirrors vascular changes occurring elsewhere in the body, including the brain.
“OCTA allows us to visualize the blood vessels to a very granular detail, to a capillary level, and noninvasively, whereas to perform a brain angiography, you need to subject the patient to an invasive procedure with intravenous dye and a much higher risk and morbidity. That is why there is a lot of emphasis on using OCTA as a screening tool initially, and potentially as a diagnostic tool in the future, to detect vascular as well as neurodegenerative conditions. And there’s a whole spectrum of neurodegenerative disorders which can be looked at,” Grewal said.
The iMIND research team at Duke University, led by Sharon Fekrat, MD, and Grewal, is involved in several studies in collaboration with neurologists in which OCT, OCTA and widefield fundus photography are used to train AI systems to identify novel biomarkers for the early diagnosis of a number of neurodegenerative conditions, including Alzheimer’s, Parkinson’s disease and other forms of cognitive impairment.
“The project started about 5 years ago, and together with my colleague Dr. Sharon Fekrat, we are working on several different lines of study,” he said.
Specifically for Parkinson’s, in 69 patients compared with 137 control participants, thinning of the retina and choroid was noted, as well as a significant reduction in vessel density.
“The branching pattern and the thickness of the blood vessels were reduced in many patients. This is the largest data set of Parkinson’s that showed these changes and clearly demonstrated significant differences from age- and sex-matched controls,” Grewal said.
The next step will be to detect how fast these changes progress over time and how these rates of change compare with normal aging.
“As we grow older, the brain shrinks and the retina thins out, and we are trying to determine if patients who have Parkinson’s or Alzheimer’s lose that tissue faster,” Grewal said.
A need for standardization
Retinal photography, OCT and OCTA enable imaging at unprecedented resolutions. However, it is still difficult to reliably obtain large quantities of good quality data across a large volume of patients and process them accurately, Grewal said.
“With all of these technologies, the more granular you get, the more susceptible they are to artifacts and quality issues. OCTA imaging is particularly susceptible to motion artifact, and almost 20% to 25% of the scans that we obtain are not reliable. This is because we are trying to image a population that is prevalently old, that may have cognitive issues, and it is hard for them to focus for the scan. The other big challenge is to get reliable software that can process the images to provide meaningful information with minimal noise,” Grewal said.
Another problem is the lack of standardized image acquisition protocols between multiple platforms and inconsistency in the reporting metrics across devices.
“If there’s a group in Europe that has 100 patients, and we have 100 patients, and we want to combine the data to increase the sample size, we cannot do that if the images are captured on different platforms. If measurements are not directly comparable to each other, we cannot aggregate information and build up a robust enough data set,” Grewal said.
AI models can be trained to recognize artifacts and other quality issues, filtering out images that are not usable, but a standardized framework for OCT and OCTA scans across the different platforms will be necessary to accelerate progress in the field.
“The research community needs to reach consensus on formats, values and parameters for the storage and analysis of imaging data to allow sharing them across the different research groups. This is critical for building up the large data sets we need in order to drive this field forward,” Grewal said.
Fluorescence methods
Amyloid beta plaques in the brain are a hallmark of Alzheimer’s disease, detectable by invasive contrast-enhanced methods. Researchers have been looking at the retina as a potentially suitable surrogate tissue to detect amyloid deposits through fluorescence methods.
“Curcumin is a food additive with fluorescent properties that binds to amyloid beta in the retina, providing fluorescent labeling and quantification of the retinal expression of the protein,” M. Francesca Cordeiro, MD, PhD, said.
Published studies with curcumin have predominantly come from Cedars-Sinai in Los Angeles, but oral administration required large quantities of tablets, leading to variable absorption. With her group at University College London, Cordeiro has been researching nasal and topical administration of curcumin to produce fluorescence in a more reproducible manner.
“The beauty of the topical eye drops and, more recently, our nasal version is that it is much better absorbed and much less of a burden. We have not tested this on patients yet, but this is what we are going to do along the line,” she said.
Another potentially promising method is DARC (detection of apoptosing retinal cells), a retinal imaging technology that uses fluorescently labeled annexin A5 (ANX776) to identify stressed and apoptotic cells in the retina. Because cell death is implicated in the pathogenesis of neurodegenerative disorders, including Alzheimer’s and Parkinson’s, DARC could offer a unique opportunity to track neuronal apoptotic changes over time.
“Since DARC also highlights stressed cells, it gives us a window of opportunity to modify cell response, preventing progression along the path of cell death. In other words, we could detect the disease earlier, when there is still reversibility,” Cordeiro said.
The evidence so far for both Parkinson’s and Alzheimer’s is predominantly in preclinical models. In Parkinson’s, retinal assessment with DARC was compared with OCT imaging and showed the ability to predict OCT changes happening at later stages. In addition, deep learning programs have been developed to use AI in the analysis of the DARC images to provide fast, automated cell count and detection of morphological changes.
Several possible markers could eventually be used and combined to overcome sensitivity and specificity issues and the problem of false positives.
“We don’t need to use just one marker. We could start off with DARC and then go on to curcumin and other types of assessment,” Cordeiro said.
A growing interest
The interest in identifying retinal imaging biomarkers of systemic disease has grown exponentially in recent years. As noted in a paper by Snyder and colleagues, a PubMed search using the terms “retina” and “Alzheimer” produced zero matches in 2000 and only seven in 2001, but 1,283 in 2019.
CVD accounted for 18 million deaths in 2019, with 85% due to heart attack or stroke, according to WHO’s statistics. Dementia, on the other hand, currently affects more than 55 million people worldwide, and nearly 10 million new cases are reported each year. Developing and validating a reliable, simple and cost-effective tool to identify people at risk for CVD and dementia would help reduce the burden of these diseases, allowing for prevention, early diagnosis and proactive treatment.
However, PET neuroimaging and cerebrospinal fluid testing are limited in their availability, have high costs, can be invasive and are time-consuming. These barriers hamper widespread population-level use.
“That is why there is a lot of emphasis now on using retinal imaging as a screening tool initially and potentially in the future as a diagnostic tool to detect vascular as well as neurodegenerative conditions,” Grewal said.
“I envisage that retina scans could become a primary screening method, followed by confirmatory assessment by cardiac CT scan in at-risk cases. And everyone going to the cardiologist could have a primary assessment by an ophthalmologist or optometrist. In other parts of the world where expensive cardiovascular assessment is not affordable, like in developing countries, retinal imaging would be an easier and more accessible way of screening for CVD,” Wong said.
The challenges
While interest is high, there are many regulatory and practical issues related to payment, insurance, interdisciplinary management and acceptance that need to be resolved.
“Some of these are what I call the painful journey of translating a very interesting concept backed by very good data and study outcomes into practical clinical application,” Wong said.
Collaboration across specialties is a crucial point, but currently no cardiologist or neurologist attends ophthalmology meetings and vice versa, so the network that could potentially lead to effective collaboration is limited.
“Interdisciplinary meetings that allow building up collaborative networks will be a very important step in making sure this technology advances,” Wong said.
Building large data sets while respecting privacy-sensitive information about individuals is another challenge. To create the database without direct consent from the patients involved, Keane and his group went through lengthy procedures for approval from the ethic committees and invoked a special legal provision, Section 251 approval, which empowers government health officials to give consent on behalf of patients for justified health research purposes. However, this remains a crucial issue to be addressed globally.
“Deep learning has powered all the advances in foreign language translation, image recognition, demographic prediction, self-driving cars and much more. Maybe we could soon have a world where we go to our family doctor, to an optometrist or to a local pharmacy and have retinal imaging performed. And the retinal imaging might be able to tell or even predict our health status and signal the need for specific specialized visits,” Keane said.
- References:
- Cardiovascular diseases. https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds). Published June 11, 2021.
- Dementia. https://www.who.int/en/news-room/fact-sheets/detail/dementia. Published Sept. 2, 2021.
- den Haan J, et al. Acta Neuropathol Commun. 2018;doi:10.1186/s40478-018-0577-2.
- Dumitrascu OM, et al. Alzheimers Dement (Amst). 2020;doi:10.1002/dad2.12109.
- Javaid FZ, et al. Front Neurol. 2016;doi:10.3389/fneur.2016.00055.
- McGeechan K, et al. Ann Intern Med. 2009;doi:10.7326/0003-4819-151-6-200909150-00005.
- Mutlu U, et al. JAMA Neurol. 2018;doi:10.1001/jamaneurol.2018.1563.
- Poplin R, et al. Nat Biomed Eng. 2018;doi:10.1038/s41551-018-0195-0.
- Snyder PJ, et al. Alzheimers Dement. 2021;doi:10.1002/alz.12179.
- For more information:
- M. Francesca Cordeiro, MD, PhD, can be reached at UCL Institute of Ophthalmology, Bath Street, London EC1V 9EL, United Kingdom; email: m.cordeiro@ucl.ac.uk.
- Dilraj Grewal, MD, can be reached at Duke Eye Center, 2351 Erwin Road, Durham, NC 27705-4699; email: dilraj.grewal@duke.edu.
- Pearse A. Keane, MD, FRCOphth, can be reached at NIHR Biomedical Research Centre for Ophthalmology, Moorfields Eye Hospital NHS Foundation Trust, 11-43 Bath Street, London, EC1V 9EL, United Kingdom; email: pearse.keane1@nhs.net.
- Tien Yin Wong, MD, PhD, can be reached at Singapore Eye Research Institute, Singapore National Eye Centre, 11 Third Hospital Avenue, Singapore 168751; email: ophwty@nus.edu.sg.
Click here to read the Point/Counter to this Cover Story.
Collapse
Read more about
Click Here to Manage Email Alerts