summary: Researchers have developed a deep learning AI tool that can automate the diagnosis of retinopathy of prematurity (ROP), a leading cause of childhood blindness.
This tool has been found to be as effective as senior pediatric ophthalmologists in distinguishing between normal retinal images and those with ROP, which can lead to blindness. Researchers hope the tool will improve access to care in underserved areas and prevent blindness for thousands of newborns around the world.
Important facts:
- Researchers have created a deep learning AI tool that can diagnose retinopathy of prematurity (ROP), a cause of childhood blindness.
- This AI tool was trained on over 7,400 neonatal eye images and was as effective as senior pediatric ophthalmologists in identifying ROP.
- As ROP becomes more common and the lack of proper infrastructure for care in some areas, AI tools may help prevent blindness in premature babies.
sauce: UCL
The team developed a deep learning AI model. This model can identify infants at risk for ROP, which can lead to blindness if left untreated, in many areas with limited neonatal services and few trained ophthalmologists. We hope that this technology can improve access to screening.
The study, by an international team of scientists and clinicians from the United Kingdom, Brazil, Egypt and the United States, was conducted at the National Institute for Health and Care Research (NIHR) Biomedical Research Center at the Moorfields Eye Hospital NHS Foundation Trust and at the UCL Institute of Ophthalmology. I am supported.is published in lancet digital health.
Lead author Konstantinos Varaskas, Ph.D., Director of Moorfields Eye Reading Center and Clinical AI Lab, Moorfields Eye Hospital and Associate Professor at UCL Eye Research Institute, said: and is now the leading cause of childhood blindness in middle-income countries and the United States. ”
“In sub-Saharan Africa, 30% of newborns have some degree of ROP, and although treatment is now readily available, it can lead to blindness if not detected and treated early. This is often due to a shortage of eye care professionals, but given that it is detectable and treatable, no child should go blind from ROP.”
“As it becomes more common, many communities do not have sufficiently trained ophthalmologists to screen every child at risk. We hope to improve access to care in underserved areas and prevent blindness for thousands of newborns around the world.”
ROP is a condition that primarily affects premature babies, where abnormal blood vessels grow in the retina. The retina is a thin layer of nerve cells behind the eye that converts light into signals that the brain can perceive. These blood vessels can leak or bleed, damaging the retina and leading to retinal detachment.
Milder ROP does not require treatment, only monitoring, while more acute cases require prompt treatment. An estimated 50,000 children worldwide are blind because of it.
Symptoms of ROP are invisible to the naked eye. That means the only way to identify the condition is to monitor an at-risk infant with an eye exam. Without an adequate infrastructure for comprehensive antenatal and postnatal care, narrow windows for screening and treatment may be missed, potentially leading to preventable blindness.
The UCL-Moorfields team developed a deep learning AI model for screening ROP. The model was trained on a sample of 7,414 images of his 1,370 neonatal eyes admitted to Homerton Hospital in London and evaluated for his ROP by an ophthalmologist.
The hospital serves an ethnically and socioeconomically diverse community. This is important because ROP can vary by ethnic group. So the tools were trained to work safely across different ethnic groups and benefit everyone.
Tool performance was evaluated on another 200 images and compared to senior ophthalmologist ratings.
The researchers further validated the tool by using it on datasets provided by the United States, Brazil and Egypt.
This AI tool proved to be as effective as senior pediatric ophthalmologists at distinguishing between normal retinal images and those with ROP, which can lead to blindness.
The tool was optimized for the UK population, but the researchers promise it has proven effective on other continents, and could be further optimized for other environments. The tool is developed as a no-code deep learning platform. This means you can optimize with new settings without any coding experience.
Lead author Siegfried Wagner, Ph.D., UCL Institute of Ophthalmology and Moorfields Eye Hospital, said: We are currently validating the tool further at multiple hospitals in the UK to learn how people interact with the AI output and understand how the tool can be incorporated into a real-world clinical setting. and ”
“We hope that this tool will allow trained nurses to capture images that can be evaluated by AI tools, allowing ophthalmologists to make treatment referrals without manually reviewing scans.” I have.”
“AI tools are particularly useful in ophthalmology, which relies heavily on manual interpretation and analysis of scans for detection and monitoring. This is where AI becomes a game-changer in the field, helping to improve vision-restoring treatments.”
About this Artificial Intelligence Research News
author: press office
sauce: UCL
contact: Press Office – UCL
image: Image credited to Neuroscience News
Original research: open access.
Siegfried K Wagner et al lancet digital health
overview
Development and international validation of a custom-designed code-free deep learning model for detection of plus disease in retinopathy of prematurity: a retrospective study
Background
Retinopathy of prematurity (ROP), the leading cause of childhood blindness, is diagnosed by interval screening by a pediatric ophthalmologist. However, the increasing survival of premature infants, coupled with the lack of available specialists, has raised concerns about the sustainability of this approach. We developed a bespoke, code-free, deep-learning-based classifier of his ROP-characteristic plus disease in an ethnically diverse population in London, UK, and analyzed the ethnic, geographic, and social characteristics of four countries. We aimed to validate them externally in economically diverse populations. three continents. Code-free Deep His Learning does not rely on the availability of professionally trained data scientists, so it is of particular potential benefit in low-resource healthcare environments.
method
This retrospective cohort study used retinal imaging of 1370 neonates admitted to the Neonatal Ward of Homerton University Hospital NHS Foundation Trust, London, UK between 2008 and 2018. Images were obtained from a Retcam version 2 device (Natus Medical, Pleasanton, CA, USA) All babies born <32 weeks of gestation or weighed <1501 g at birth. Each image was graded by her two junior ophthalmologists and discrepancies were adjudicated by a senior pediatric ophthalmologist.
A custom-built, code-free deep learning model (CFDL) was developed to identify healthy, pre-plus disease, and plus disease. Performance was assessed internally on 200 images with a majority vote of his 3 senior pediatric ophthalmologists as a reference standard. External validation was performed on 338 retinal images from his four separate datasets from the United States, Brazil, and Egypt using images derived from Retcam and 3nethra neo devices (Forus Health, Bangalore, India). Done.
findings
Of the 7414 retinal images in the original dataset, 6141 images were used in the final development dataset. For discrimination between healthy and pre-plus or plus disease, the custom model had an area under the curve (AUC) of 0.986 (95% CI 0.973–0.996) and the CFDL model had an AUC of 0. 989 (0 979–0 997) internal test set. Both models generalized well to external validation test sets acquired using Retcam and discriminated healthy from pre-plus or plus disease (tailored range 0.975–1.000 and CFDL range 0.969–1.000). 0.995). The CFDL model is based on the US dataset (CFDL 0.808 [95% CI 0·671–0·909, bespoke 0·942 [0·892–0·982]], p=0.0070). 3nethra neo imaging device (CFDL 0 865 [0·742–0·965] Made to Order 0・891 [0·783–0·977]).
interpretation
Both custom and CFDL models gave similar performance to senior pediatric ophthalmologists for discriminating healthy retinal images from those with pre-plus or plus disease features. However, considering minority classes, the CFDL model may not generalize well. Care should be taken when testing data acquired using a different imaging device than the one used for the development dataset. Our study justifies further validation of the positive disease classifier in ROP screening and supports the potential role of code-free approaches in helping prevent blindness in vulnerable neonates.
fundraising
National Institutes of Health Biomedical Research Center based at Moorfields Eye Hospital NHS Foundation Trust and University College London Institute of Ophthalmology.
