AI to the rescue in glaucoma? Doctors at APAO-AIOC 2025 explain how advancements in the disease are outpacing the doctors trying to treat it.
Artificial intelligence (AI) and advanced imaging are shaking up the glaucoma landscape—and not a moment too soon. This was the central message from a Day 4 symposium at the 40th Congress of the Asia-Pacific Academy of Ophthalmology (APAO 2025)—held in conjunction with the 83rd Annual Conference of the All India Ophthalmological Society (AIOC 2025) in New Delhi.
The growing crisis in glaucoma care
Dr. Benjamin Xu (USA) kicked things off by spotlighting a harsh reality: glaucoma is on the rise, and the workforce to treat it isn’t keeping pace.
“The prevalence of glaucoma has been dramatically rising worldwide primarily due to aging of the world population. At the same time, in many countries, the size of the ophthalmology workforce is declining due to physician burnout and early retirement,” he explained.
That widening gap is especially evident in glaucoma subspecialty care. Dr. Xu presented data showing that while the United States has increased the number of glaucoma fellowship positions, the number of filled spots has remained flat—a troubling mismatch between supply and demand.1,2
AI tackles workflow woes
Dr. Xu pinpointed two major bottlenecks where AI could make a real difference. First up: unnecessary referrals. At Kaiser Permanente, his team found that only 8.2% of patients aged 18 to 40 referred for glaucoma evaluation actually received a diagnosis within two years.3
To streamline this, Dr. Xu and colleagues launched the ATLAS (AI and Teleophthalmology in Los Angeles) initiative and trained a deep learning model to detect referable glaucoma from fundus photos. The result?
“We compared the performance of this algorithm, and we found that it matched or exceeded the performance of all human graders,” he reported. “Among human graders, there’s a wide range of sensitivity and specificity, which highlights the weakness of using human graders in this type of screening program.”
The second issue: gonioscopy. Despite being a key part of initial glaucoma evaluations, it’s often skipped—possibly because it disrupts the clinic workflow. Dr. Xu’s analysis of 200,000 patient records revealed that fewer than 30% received gonioscopy within six months of evaluation, despite professional guidelines.4
To address this, his team turned to anterior segment optical coherence tomography (AS-OCT), building an algorithm that can spot angle closure with close to 85% sensitivity and specificity—and with consistent performance across diverse patient groups.
The promise (and pitfalls) of AI
Next up was Dr. Fei Li from Zhongshan Ophthalmic Center (China), who introduced iGlaucoma, a multimodal AI diagnostic platform.
While optimistic about AI’s future, he was refreshingly candid while listing its current limitations. “First, the reliance on large amounts of data for training; second, poor interpretability or the ‘black box’ problem; and third, inadaptability to the real world.”
To tackle the data dilemma, Dr. Li pointed to federated learning—a privacy-preserving method of data sharing that enables validation across sites without compromising security.
And to address the ‘black box’ problem, his team uses Grad-CAM (Gradient-weighted Class Activation Map) to generate heat maps that show where the AI is focusing.
“The heated regions align with human knowledge in glaucoma diagnosis,” Dr. Li explained, though he noted that these tools “only provide insight into the discriminative regions of the image and cannot reveal the internal mechanisms of the model.”
Looking ahead, he suggested that large language models (LLMs) might bridge the interpretability gap. “These models can interact with the users. Somehow people may know how they work. For instance, they could tell you the description of a fundus photo, the boundary of the cup and disc, how they estimate the CD [cup-to-disc] ratio, etcetera.”
Seeing beyond the surface with ROTA
Prof. Christopher Leung (Hong Kong) introduced an exciting new tool: ROTA (Regular Nerve Fiber Layer Optical Texture Analysis). This imaging technique combines thickness and reflectance information from the macular nerve fiber layer to better spot defects—especially in tricky cases.
“ROTA allows us to visualize the axonal fiber bundles of the retina, enabling us to identify retinal nerve fiber defects when conventional mechanical instruments like OCT…fail to identify,” he explained.
Using clinical examples, Prof. Leung demonstrated how ROTA outperformed traditional OCT and fundus photography—particularly in myopic eyes, where standard assessments often fall short.
“We validated the performance of ROTA using visual field testing as the recommended standard,” he added. “What we found is that ROTA outperforms conventional OCT analysis in detecting glaucoma.”
Genetic, risk and the AI advantage
Prof. Toru Nakazawa (Japan) took a broader view, discussing how AI could enhance genetic risk stratification. He highlighted the importance of polygenic risk scores (PRS) in stratifying glaucoma risk.
“High PRS of glaucoma has been shown to be associated with more severe phenotype, a higher age onset and faster progression,” he noted, citing recent research.
Prof. Nakazawa believes that a great deal of AI’s potential in glaucoma lies in integrating diverse data sources for comprehensive risk assessment. “AI multimodal analysis is useful in glaucoma because it allows for polygenic risk assessment. Knowing the key risk factors for each glaucoma patient allows us to provide more suitable treatment and obtain better prognosis,” he concluded.
Decoding NTG with AI
Normal tension glaucoma (NTG) can be tough to pin down—but AI may lend a hand. Prof. Ki Ho Park (South Korea) shared how deep learning has boosted the detection of disc hemorrhage, a subtle but telling NTG marker.
“The use of enhanced optic disc photos led to a significant improvement in the accuracy of detecting disc hemorrhage,” he reported.
Prof. Park’s team also explored explainability, using adversarial examples to better understand model decisions. “Adversarial examples increased explainability over a graph- and a heat map-based explanation method,” he explained.
Even more intriguing, they found that combining fundus images with clinical data helped predict whether—and when—NTG suspects would convert to full-blown disease.5
Where do we go from here?
This session served up a compelling mix of optimism and realism. AI and imaging are clearly on the rise, but success in the clinic will hinge on real-world validation, better interpretability, and strategic integration into workflows.
As Dr. Xu put it, “AI is a transformative technology that promises to change the way that we deliver eye care. However, AI is just a tool, and it is up to clinicians and researchers to identify barriers to care efficiency and properly apply these AI tools to address these issues. We need more real-world implementation studies before this can become widely useful.”
Dr. Li drew an intriguing parallel to a similar trend from ophthalmology’s past. “AI is like genetics was 20 years ago—a burgeoning tool yet distant from clinical application, requiring exploration.”
Still, with AI, advanced imaging and multimodal analysis all gaining momentum, the future of glaucoma is looking sharper—and smarter—by the day.
References
- Association of University Professors of Ophthalmology. Ophthalmology Fellowship Match Statistics & Demographics Dec 2021. AUPO. Mar 20, 2022. Available at: https://aupo.org/news/2022-03/ophthalmology-fellowship-match-statistics-demographics-dec-2021. Accessed on April 6, 2025.
- Burton E, Arsiwala LT, Johnson TV, et al. Applicant characteristics associated with glaucoma fellowship match from 2010 to 2017. Ophthalmol Glaucoma. 2022 Mar-Apr;5(2):233-240.
- Frediani T, Yoo K, Cho A, et al. Outcomes of glaucoma referrals in adults aged 18 to 40 years. JAMA Netw Open. 2025 Feb 3;8(2):e2457843.
- Lee JH, Yoo K, Xu B, et al. Patterns and disparities in recorded gonioscopy during initial glaucoma evaluations in the United States. Am J Ophthalmol. 2024 Aug;264:90-98.
- Ha A, Sun S, Park KH, et al. Deep-learning-based prediction of glaucoma conversion in normotensive glaucoma suspects. Br J Ophthalmol. 2024 Jun 20;108(7):927-932.
Editor’s Note: Reporting for this story took place during the 40th Congress of the Asia-Pacific Academy of Ophthalmology (APAO 2025), being held in conjunction with the 83rd Annual Conference of the All India Ophthalmological Society (AIOC 2025) from 3-6 April in New Delhi, India.