Article
News
ARVO 2024: AI may open new frontiers in vision research
Author(s):
Four poster presentations at the Association for Research in Vision and Ophthalmology’s (ARVO) 2024 Annual Meeting in Seattle, Washington, demonstrate the possibilities.
(Image credit: Adobe Stock/kellyvandellen)
A quartet of presentations this week at the Association for Research in Vision and Ophthalmology’s (ARVO) 2024 Annual Meeting in Seattle, Washington.
According to a news release, the presentations demonstrated the benefits of involving artificial intelligence (AI) in vision research and global health. They showed the advantages of having AI systems to support scientists, clinics and patients.
Promising new anterior eye imaging tool
In an abstract titled Optical transmission tomography: Technology and capabilities of a novel anterior eye imaging tool, a team in France developed a new imaging tool for anterior eye imaging, optical transmission tomography (OTT). OTT was inspired by two scientific approaches, phase-contrast microscopy and asymmetric retroillumination microscopy. According to a news release, most imaging methods rely on reflection, whereas OTT uses back illumination from the eye’s fundus and a special interference technique to highlight the front part of the eye with great clarity.
This approach provides a new perspective on the cellular organization of the eye. When it was used to study two young and healthy individuals, OTT’s viewing area was 25 times larger than that of the specular and confocal microscopies and three times larger than that of an advanced research system such as curved-field optical coherence tomography (OCT).1
Viacheslav Mazlin, PhD, the lead scientist, addressed the findings.
“Micro-resolution across the macro-scale in non-contact OTT provides a new look into the cellular changes involved in anterior eye conditions,” Mazlin said in the reease. “OTT holds the potential to improve our understanding of dry eye disease and improve outcomes of refractive/cataract surgeries undertaken by millions of people each year.”
Enhancing inherited retinal disease datasets
In a presentation titled Creating the world’s largest dataset of segmented Inherited Retinal Disease features by bootstrapping manual annotations with AI, researchers set out to enhance inherited retinal disease databases.
To understand and diagnose inherited retinal diseases (IRDs), phenotype-genotype recognition is used. Researchers noted in the news release this requires looking at both visible signs in the retina, phenotype, and the genetic makeup, genotype, to understand these diseases. Images of the retina are analyzed manually using a variety of techniques to identify these features, wn be a laborious and time-consuming subjective process, usually only executed by trained specialists in IRDs.1
William Woof, PhD, the lead data scientist of the Eye2Gene team from various institutions in the United Kingdom, detailed the research in the news release.
“We developed a set of AI algorithms to automatically identify features within retinal scans of IRD patients with a known gene diagnosis and used them to create a large IRD dataset for research, he said.
A subset of manually segmented data from optical coherence tomography (OCT) and fundus autofluorescence (FAF) scans were used to train the AI models. Then, these models were used to analyze the remaining scans in the entire dataset, providing them with a large database.1
Woof noted the researchers found that using AI was effective in analyzing big datasets of retinal images.
“This dataset will help deepen our understanding of the phenotype-genotype patterns in various IRDs and, along with our AI models, lower the barrier to entry for quantifying the potential impact of new treatments,” Woof concluded in the news release.
Ensuring equity in AI’s role in diabetic retinopathy detection
In a presentation titled Evaluation of equity in performance of Artificial Intelligence for diabetic retinopathy (DR) detection, researchers set out to ensure equity as AI is used in detecting diabetic retinopathy.
With the increasing prevalence of diabetes, the expenses and workload related to screening for diabetic eye diseases follow suit. Prior studies have demonstrated that automated technology for detecting diabetic eye diseases from eye images can identify images with DR as accurately as human graders. According to the news release, this would offer the potential for substantial workload reduction.
A team of United Kingdom researchers led by Alicja Rudnicka, PhD, looked at automated retinal image analysis systems (ARIAS) in 200,000 screening visits, approximately 1.2 million images.
They discovered that they performed consistently well for individuals with diabetes in identifying moderate-to-severe diabetic eye disease. This finding suggests that ARIAS could serve as an effective tool for triaging individuals into high-risk groups for human grading and low-risk groups not needing human grading.1
Rudnicka pointed out that AI algorithms accurately detect diabetic eye disease, potentially solving a global problem.
“Medical devices need to have equitable outcomes across sociodemographic groups, particularly by sex, age, level of deprivation and ethnicity,” Rudnicka said in the news release. “Our independent study measures and describes the algorithmic fairness of multiple commercially available algorithms on real-world screening data…the largest and most diverse study of its kind — thus contributing to public and clinical trust and, we hope, widespread implementation of this transformative technology."
This study was funded by the National Institute for Health and Care Research (NIHCR).
Evaluating a novel AI system in managing and screening diabetic retinopathy
The management and regular screening of diabetic retinopathy (DR) performed by primary care physicians (PCPs) continues to be a global health challenge, especially in low-resource areas.
In a paper titled An Integrated Image-based Deep Learning and Language Models for Diabetic Retinopathy: A Multi-Stage Development, Testing and Prospective Comparative Study, an international research team led by Tien Y. Wong, MD, PhD, FARVO, created a new deep learning-large language model system (DeepDR-LLM) to assist PCPs in improving DR screening and diabetes care.1
The researchers noted in the news release the AI system utilized fundus images and integrated deep learning for image analysis with the capabilities of large language models (LLMs). The LLM module was developed using 371,763 management suggestions from 267,730 patients. A prospective study was performed on 769 patients to determine its impact.
According to the news release, they then examined the “adherence to management recommendations” among patients who received care from PCPs without assistance and those who received care from PCPs supported by DeepDR-LLM. The study showed that newly diagnosed patients under the care of PCPs assisted by DeepDR-LLM expressed better self-management tactics, were more likely to follow DR referral, and showed improvement in both the quality and the level of empathy in management recommendations.1
“We have developed a novel AI system for diabetic retinopathy screening that combines an automatic assessment of DR status from the fundus image, with a LLM-guided recommendation to individual patients,” Wong explained, “removing the need for specialized ophthalmology services for DR screening, particularly in low-income countries.”
