advertisement
Editors Selection IGR 25-1
Artificial Intelligence: AI predicts RNFL thickness from disc photos to flag glaucoma risk
Comment by Xiulan Zhang & Zefeng Yang on:
127685 Predicting Retinal Nerve Fiber Layer Thickness From Ocular Hypertension Treatment Study Optic Disc Photographs, Liu JC; Jammal AA; Scherer R et al. et al., JAMA ophthalmology, 2025; 143: 652-659
Effective management of ocular hypertension requires precise risk stratification to prevent conversion to glaucoma. While the landmark Ocular Hypertension Treatment Study (OHTS) identified key risk factors, its reliance on subjective optic disc grading introduced significant inter-observer variability. Artificial intelligence (AI), especially the “Machine-to-Machine" (M2M) deep learning models”, now enables the mapping of features from one modality to another. The AI-driven prediction of retinal nerve fiber layer (RNFL) thickness from optic disc photographs may help to determine risk for development of primary open-angle glaucoma (POAG) in patients with ocular hypertension. In this diagnostic study by Liu et al., researchers applied an M2M deep learning model, trained with 32,820 pairs of fundus photographs and the associated OCT scans, to 64,755 optic disc photographs from 1,444 participants in the OHTS cohort. The model successfully stratified risk, showing that baseline predicted RNFL was significantly thinner in converters (94.1 µm) compared to non-converters (97.1 µm), with a hazard ratio (HR) of 1.83 per 10-µm thinning. Notably, the longitudinal rate of change proved even more predictive: converters demonstrated an RNFL thinning rate nearly double that of non-converters (-0.42 µm/year vs -0.25 µm/year). A 1-µm/year faster loss was associated with a 6-fold increase in conversion risk. This study effectively validates M2M-derived RNFL as a potent biomarker for risk stratification. However, several technical and clinical nuances warrant future refinement. First, because this is a secondary analysis of the historical OHTS dataset, lacking prospective clinical validation, these predicted RNFL lack validation against ground-truth OCT measurements. Second, the model's mean absolute error of 7.4 μm deriving slopes from these estimated values risks amplifying measurement noise1, which may fail to accurately reflect the true slope. Finally, the reliance on serial fundus photography to calculate slopes limits its utility in opportunistic screening, where clinicians prioritize immediate prediction at the first encounter over long-term monitoring2-3. Future research should aim to overcome these barriers by integrating AI-generated structural metric with raw imaging features or functional modalities. Such multimodal integration could mitigate estimation noise and transform the current risk model into a precise, personalized prediction engine for individual patient care.
References
- Medeiros FA, Jammal AA, Thompson AC. From Machine to Machine: An OCT-Trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs. Ophthalmology. 2019 Apr;126(4):513-521. https://doi.org/10.1016/j.ophtha.2018.12.033 /li>
- Li F, Su Y, Lin F, Li Z, Song Y, Nie S, Xu J, Chen L, Chen S, Li H, Xue K, Che H, Chen Z, Yang B, Zhang H, Ge M, Zhong W, Yang C, Chen L, Wang F, Jia Y, Li W, Wu Y, Li Y, Gao Y, Zhou Y, Zhang K, Zhang X. A deep-learning system predicts glaucoma incidence and progression using retinal photographs. J Clin Invest. 2022 Jun 1;132(11):e157968. https://doi.org/10.1172/JCI157968
- Thakur A, Goldbaum M, Yousefi S. Predicting Glaucoma before Onset Using Deep Learning. Ophthalmol Glaucoma. 2020 Jul-Aug;3(4):262-268." https://doi.org/10.1016/j.ogla.2020.04.012
Comments
The comment section on the IGR website is restricted to WGA#One members only. Please log-in through your WGA#One account to continue.Log-in through WGA#One
