Multi-Comparison of Different Ocular Imaging Modality-based Deep Learning Models for Visually Significant Cataract...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Age-related cataract is a primary cause of vision impairment globally.
• Deep learning (DL) algorithms are being developed for automated cataract analysis using various imaging techniques.
• Comparative performance of DL models across different ocular imaging modalities is not well-established due to a lack of standardized datasets.

Purpose of the Study:

• To evaluate and compare the performance of deep learning (DL) algorithms for detecting visually significant cataract (VSC) across different ocular imaging modalities.
• To assess single-modality DL models (retinal, slit beam, diffuse anterior segment) and ensemble models.

Main Methods:

• Developed three single-modality DL models and four ensemble models for VSC detection.
• Utilized large datasets (Singapore Malay Eye Study, SINDI, SCES) for training and external validation.
• Defined VSC based on the modified Wisconsin cataract grading system and best-corrected visual acuity <20/60.

Main Results:

• The retinal imaging DL model achieved the highest Area Under the Receiver Operating Characteristic Curve (AUC) in internal testing (97.0%) and external testing.
• The retinal model outperformed slit beam (93.4%) and diffuse anterior segment (94.4%) models in internal testing.
• Retinal model demonstrated reasonable performance in nonmydriatic retinal photos (AUC, 89.8%).

Conclusions:

• The retinal imaging DL model is a promising tool for detecting VSC, surpassing other tested modalities.
• Retinal photography's routine use in diabetic retinopathy screening allows for cost-effective, opportunistic cataract screening.
• This approach can integrate cataract detection into existing eye screening programs.