Cataract is a general term for diseases in which part or all of the lens becomes cloudy. An estimated 52.6 million people worldwide have moderate to severe visual impairment, and it is considered a leading cause of treatable blindness. Cataract surgery is one of the most cost-effective medical interventions, but in low- and middle-income countries, limited resources and long waiting times are barriers to treatment.
In recent years, digital technologies, particularly artificial intelligence (AI), have been applied to multiple stages of cataract care. The application areas of AI technology can be broadly divided into the following three categories.
With the aging population, the demand for cataract surgery is increasing, while the supply of ophthalmic care is not keeping up. AI technology is attracting attention as a means to bridge this supply-demand gap.
With the increase in the elderly population, the demand for cataract surgery is rapidly rising, while the supply of ophthalmic care has not expanded sufficiently. Particularly in low- and middle-income countries and rural areas, there are many undiagnosed cataracts, and AI-assisted remote diagnosis and screening have the potential to reduce disparities in healthcare access.
Cataracts are primarily caused by aging, with prevalence rates including early opacities reaching approximately 45% in people in their 50s, 75% in their 60s, 85% in their 70s, and 100% in those aged 80 and older. The main types of opacities are cortical, nuclear, and posterior subcapsular, with cortical cataracts being the most common in Japanese people.
For cataract severity classification, LOCS III (Lens Opacities Classification System III), WHO classification, and Emery-Little classification are used. The Emery-Little classification evaluates nuclear color in five grades and is widely used to estimate surgical difficulty.
In the early stages of cataracts, there are often no subjective symptoms. As the condition progresses, the following symptoms appear.
No medication exists to clear cataracts. For cases with reduced visual function, phacoemulsification (PEA) and intraocular lens (IOL) implantation are the standard treatment. Advances in surgical instruments allow surgery through a small incision of about 2 mm, enabling early return to daily life. Recently, femtosecond laser-assisted cataract surgery is also used clinically.
Cataracts are currently diagnosed clinically by ophthalmologists using a slit-lamp microscope, requiring an in-person examination. However, in developing countries and rural areas, access is difficult, making undiagnosed cataracts a major problem.
AI-assisted remote diagnostic platforms can reduce these access barriers. Rapid diagnosis is especially critical in pediatric cataracts to prevent irreversible amblyopia.
Slit-Lamp Photo-Based
Wu et al.’s model: Trained on approximately 38,000 slit-lamp photos. Achieved over 95% sensitivity and specificity for cataract detection and severity classification. Reported about 80% sensitivity and specificity for three-grade nuclear hardness assessment.
Li et al. (Visionome): Diagnosed anterior segment diseases including cataracts from slit-lamp photos. Showed accuracy of 79.47–99.22%, outperforming ophthalmologists with one year of clinical experience.
Fundus Photo-Based
Xu et al.’s model: CNN ensemble algorithm (AlexNet + VisualDN) using fundus photos as input. Achieved 86.2% accuracy for cataract detection and classification.
ResNet-based model (Wu et al.): Reported AUC over 0.99 for three-class discrimination among cataractous lenses, intraocular lens eyes, and normal eyes.
Wu et al. proposed a three-step model: (1) self-monitoring using a smartphone, (2) AI diagnosis using anterior segment photographs, and (3) teleconsultation with an ophthalmologist via a cloud platform. This mechanism is said to increase the population covered by a single ophthalmologist tenfold. AI diagnosis in ophthalmology has advanced in retinal diseases and glaucoma using fundus photographs and posterior segment OCT since the 2016 publication on diabetic retinopathy screening, but in recent years, its application to anterior segment slit-lamp photographs has progressed.
Although still in the research stage, systems have been proposed in which AI analyzes anterior segment images taken with a smartphone to determine the presence and severity of cataracts. Further validation is needed for clinical application, but it may be useful for screening in remote areas.
Cataract surgery requires accurate postoperative refractive outcomes. Conventional intraocular lens formulas have insufficient predictive accuracy in eyes after refractive surgery or in patients with extreme biometric values. New-generation formulas utilizing AI are addressing this challenge.
| Formula | Features | Remarks |
|---|---|---|
| Kane formula | Theory + AI regression analysis | Consistently top 3 performance in comparative studies |
| Hill-RBF | ANN-based pattern recognition | Analyzes large refractive datasets |
| PEARL-DGS | ML + linearization of output | Adjusts for extreme biometric values |
Kane formula combines a theory-based model with regression analysis and AI components. In comparative studies, it outperforms third-generation formulas such as Barrett Universal II, Haigis, and Olsen, and consistently ranks among the top three new-generation formulas. It also shows reasonable results in extreme axial length cases.
Hill-RBF (radial basis function) is an artificial neural network-based intraocular lens calculator that analyzes a large dataset of refractive outcomes using pattern recognition and data interpolation.
PEARL-DGS formula uses machine learning modeling and output linearization to predict effective lens position and adjust for extreme biometric values.
Karmona is a data-driven intraocular lens power calculation approach that uses multiple machine learning models such as k-nearest neighbors, ANN, SVM, and random forest to predict the power.
Traditional intraocular lens formulas are based on specific theoretical models, which can lead to larger errors in eyes after refractive surgery or with extremely long or short axial lengths. AI-based formulas learn patterns from large datasets, maintaining high prediction accuracy even in these special cases.
AI is also applied in cataract surgery training, intraoperative decision-making, and postoperative analysis.
Cataract surgery consists of multiple phases (e.g., capsulorhexis, lens nucleus processing, cortex aspiration, intraocular lens insertion). Automated analysis of surgical videos using AI enables identification of each phase.
Automatic phase identification serves as the foundation for phase-specific assessment of surgical skills and real-time feedback.
Lanza et al. analyzed 1,229 cataract surgeries including 73 errors and built an AI model to detect risk factors for intraoperative complications and predict total surgical time.
Eyesi (Haag-Streit) is a commercially available ophthalmic simulation-based training system that provides intelligent education by integrating virtual reality (VR) and AI. It creates an environment where learners can acquire surgical skills before interacting with real patients.
If AI is successfully introduced into clinical cataract care, long-term benefits such as improved medical efficiency, better access, and cost reduction are expected. The benefits will be particularly significant for low-income populations.
However, the following challenges exist for practical implementation.
To date, only a limited number of algorithms have demonstrated reliability in clinical environments. Further randomized controlled trials are needed to establish the usefulness of AI systems.
Although high accuracy has been reported at the research level, there are still challenges for full-scale introduction into clinical practice. Confirming generalizability across different populations, validating effectiveness through large-scale clinical trials, and establishing ethical data management are essential. For now, practical use as a tool to assist ophthalmologists in diagnosis is realistic.
