- Retinopathy of prematurity is an eye condition that affects preterm babies and can cause visual impairment or blindness unless detected and treated during the early stages of the disease.
- Regular screening of preterm babies can help prevent these adverse outcomes, but there is a shortage of pediatric ophthalmologists, especially in low-income and middle-income countries.
- A recent study showed that an artificial intelligence (AI) model could analyze images of the retina and accurately diagnose retinopathy of prematurity in preterm babies.
- The AI model used in the study did not require coding experience and could be potentially deployed in resource-limited settings.
Severe retinopathy of prematurity can cause visual impairment and blindness in children. The condition is one of the leading cause of childhood blindness.
Although screening programs can help prevent the progression of retinopathy of prematurity, there are concerns about the scarcity of pediatric ophthalmologists to perform these screenings, especially in resource-limited settings.
Studies have shown that AI applications can accurately diagnose severe retinopathy of prematurity based on the analysis of retina images. However, the development of these AI applications requires the expertise of data scientists and expensive hardware.
A recent study published in the journal Lancet Digital Health reports that a code-free AI application that does not require coding expertise or expensive hardware could accurately detect severe retinopathy of prematurity using images obtained from an ethnically diverse dataset from the United Kingdom as well as those acquired in low-income and middle-income countries such as Brazil and Egypt.
The researchers said that this AI model could diagnose severe retinopathy of prematurity using images obtained with a device other than the one used for developing the model, albeit with a reduction in accuracy.
Although further validation is needed, the researchers said their findings indicate that code-free AI models may possess the potential for accurately diagnosing retinopathy of prematurity in resource-limited settings.
“As many as 30 percent of newborns in sub-Saharan Africa have some degree of retinopathy of prematurity and, while treatments are now readily available, it can cause blindness if not detected and treated quickly,” said Dr. Konstantinos Balaskas, a study author and an associate professor at University College London. “This is often due to a lack of eye care specialists, but, given it is detectable and treatable, no child should be going blind from retinopathy of prematurity.”
“As it becomes more common, many areas do not have enough trained ophthalmologists to screen all at-risk children,” Balaskas told Medical News Today. “We hope that our technique to automate diagnostics of retinopathy of prematurity will improve access to care in underserved areas and prevent blindness in thousands of newborns worldwide.”
Screening for retinopathy of prematurity
Retinopathy of prematurity is an eye disease that affects the retina, which forms the inner layer of the eye and is responsible for converting light into nerve impulses.
Retinopathy of prematurity is generally observed in infants born before 31 weeks of pregnancy or with a body weight of under 3 pounds.
This eye condition is caused by the abnormal growth of blood vessels in the retina. In mild retinopathy of prematurity, the changes in the blood vessels in the retina resolve on their own. In contrast, the abnormal growth of blood vessels in severe retinopathy of prematurity can cause the retina to detach, leading to blindness.
Severe retinopathy of prematurity is characterized by structural changes involving the enlargement and twisting of blood vessels in the retina, referred to as plus disease. The presence of plus disease is considered to be a marker of retinopathy that requires treatment.
Current guidelines recommend periodic screening of preterm or low birth weight infants by pediatric ophthalmologists. While there have been considerable improvements in the survival of premature infants due to technological advances and increased screening, the lack of an adequate number of pediatric ophthalmologists is an obstacle to the sustainability of this effort.
The scarcity of pediatric ophthalmologists is even more acute in lower-income and middle-income countries. Over the past decade, artificial intelligence applications have shown promise in addressing this issue, but there are a few obstacles to using this innovative approach to screening.
Use of artificial intelligence models
Ophthalmologists use images of the retina to visualize blood vessels and diagnose plus disease. Over the past decade, artificial intelligence applications have been developed that can analyze imaging data and diagnose retinopathy of prematurity as accurately as experienced ophthalmologists.
Specifically, these applications are based on deep learning, a form of artificial intelligence that simulates the process of learning that occurs in the brain. Before being deployed for diagnosing diseases, deep learning models are trained using an imaging dataset annotated or labeled by medical experts. For retinopathy of prematurity, this would involve using images that ophthalmologists have previously identified as healthy or with plus disease.
However, there are several obstacles to the direct deployment of these models for the diagnosis of plus disease in the clinic, especially in low and middle-income countries. For instance, most of these deep learning models have been optimized using data from North America and Asia.
These data are expected to underrepresent ethnic groups and those from a lower socioeconomic background. The development of retinopathy of prematurity is influenced by ethnicity, suggesting that these models may not be generalizable.
Furthermore, research groups have trained most of these AI models for the detection of plus disease using data obtained with a specific imaging device called Retcam. Imaging devices such as Retcam tend to be costly and other devices are often used in lower-income and middle-income countries.
However, the accuracy of these models has yet to be assessed on datasets obtained using other imaging devices. AI algorithms often show a decline in accuracy when deployed to analyze imaging data obtained using a different device than the one used for model development, highlighting the need to validate these models on external datasets before real-world deployment.
The deployment of these AI models is also limited by the need for expensive computer hardware and the expertise of data scientists. These resources may not be accessible to individual clinicians and even research groups, especially in lower-income and middle-income nations.
These impediments associated with customized deep learning models can be circumvented by the use of code-free deep learning applications that do not require coding expertise and have an easy-to-use interface. Moreover, code-free deep learning programs are often cloud-based, thus negating the need for costly hardware. These code-free deep learning platforms still require an annotated dataset but can be used by a clinician without coding experience.
Accuracy of the code-free deep learning model
In the present study, the researchers compared the performance of a bespoke and code-free deep learning model with experienced clinicians in diagnosing plus disease based on analyzing imaging data from different countries obtained using Retcam.
Moreover, they examined the ability of these models developed using Retcam to accurately identify plus disease using images obtained with a different device.
The researchers first developed a bespoke and code-free deep learning model using Retcam images obtained from newborns from ethnically and socioeconomically diverse backgrounds at a United Kingdom hospital. Specifically, the bespoke and code-free deep learning models were initially trained on a subset of images from these neonates and then their accuracy was evaluated on the remaining images from this dataset.
The bespoke and code-free deep learning models showed similar accuracy to senior ophthalmologists in detecting infants without or with plus disease or pre-plus disease. Pre-plus disease describes abnormalities in blood vessels similar to those seen in plus disease but are not severe enough to be diagnosed as plus disease. Detection of pre-plus disease can help initiate early treatment of retinopathy of prematurity.
The two models also showed similar high diagnostic accuracy while analyzing Retcam image datasets from the United States and two lower-income and middle-income countries – Brazil and Egypt. However, the code-free deep learning model showed lower accuracy in detecting cases with pre-plus disease than the bespoke model.
The researchers also assessed the performance of the models using a separate dataset from Egypt obtained with a different imaging device called 3nethra. Both models showed a decline in diagnostic accuracy during the analysis of this dataset obtained using 3nethra than the training or validation datasets.
These results highlight the potential of the code-free deep learning model for the diagnosis of plus disease in low- and middle-income countries where the scarcity of pediatric ophthalmologists and limited resources may hinder the regular screening of preterm infants.
“This is a clever study that shows a potentially very useful application of artificial intelligence. The authors showed that their AI program performed as well as senior eye doctors in identifying a leading cause of blindness in children by examining retinal images,” said Dr. Deepak Bhatt, MPH, the director of Mount Sinai Heart in New York.
“Machine learning and AI have moved out of science fiction to possible utility in clinical practice,” Bhatt told Medical News Today. “This study is a nice example of that. More studies like this are needed in diverse populations.”
Source: Read Full Article