The Use of Artificial Intelligence for Estimating Trabecular–Iris Angle from Slit-Lamp Images Developed Using Anterior-Segment Optical Coherence Tomography

Abstract

Background: Narrowing of the trabecular–iris angle (TIA) is a key anatomical feature of primary angle-closure disease and contributes to impaired aqueous humor outflow; however, conventional angle assessment requires specialized expertise or costly imaging equipment. This study aimed to develop an artificial intelligence model for quantitative TIA estimation from anterior-segment images acquired with a smartphone-attached slit-lamp microscope. Methods: In this retrospective study, 127,512 candidate frames were extracted from anterior-segment videos obtained from 2654 eyes. After automated diagnosability filtering, 59,412 frames suitable for TIA500 estimation were retained and assigned to the training, validation, and test datasets using patient-level grouping. A ResNet50-based model was trained to estimate inferior, nasal, and temporal TIA500 values using the corresponding measurements obtained by anterior-segment optical coherence tomography (AS-OCT) as the ground-truth reference. Model performance was evaluated by comparing the AI-predicted values with the AS-OCT measurements using mean absolute error (MAE), root mean squared error (RMSE), Spearman’s rank correlation coefficient, and the intraclass correlation coefficient (ICC). Results: Per-eye MAEs were 7.01° ± 5.69°, 7.28° ± 5.80°, and 7.90° ± 6.37° for inferior, nasal, and temporal TIA, respectively. Corresponding correlations with reference measurements were 0.683, 0.696, and 0.646, while ICCs were 0.769, 0.764, and 0.743 (all p < 0.001). Conclusions: These findings demonstrate the feasibility of estimating TIA500 from smartphone-based slit-lamp images. However, the moderate correlations with AS-OCT measurements indicate that the present model is not yet sufficiently accurate for independent clinical diagnosis or screening, and further model improvement and external validation are required.