Abstract
Objectives: To develop an algorithm to assist in the diagnosis of glaucoma with concomitant retinal vein occlusion (RVO) and to compare its diagnostic accuracy with that of ophthalmology residents and specialists. Methods: Fundus photographs of eyes with RVO and those with both RVO and glaucoma were obtained from patients who visited the University of Yamanashi Hospital. All images were preprocessed through normalization and resized to 512 × 512 pixels to ensure uniformity before model training. The diagnostic accuracy of two algorithms-the Comprehensive Fundus Disease Diagnostic Artificial Intelligence Algorithm (CD-AI) and the Glaucoma Concomitant RVO Artificial Intelligence Algorithm (RVO-GLA AI)-was evaluated. CD-AI is a clinical decision support algorithm originally developed to detect eleven common fundus diseases, including glaucoma and RVO. RVO-GLA AI is a fine-tuned version of CD-AI that is specifically adapted to detect glaucoma with or without RVO. Fine-tuning was performed using 1234 images of glaucoma, 1233 images of nonglaucomatous conditions, including RVO, and 15 images of cases with both glaucoma and RVO. The number of comorbid cases was determined empirically by gradually adding glaucomatous eyes with concomitant RVO to the training set, and 15 images provided the best balance between sensitivity and specificity. Because the available number of such cases was limited, this small sample size may have influenced the stability of the performance estimates. For the final evaluation, both algorithms and all ophthalmologists assessed the same independent test dataset comprising 66 fundus images (16 eyes with glaucoma and RVO and 50 eyes with RVO alone). The diagnostic performance of both algorithms was compared with that of three first-year ophthalmology residents and three board-certified ophthalmologists. Results: CD-AI demonstrated high diagnostic accuracy (92.5%) in eyes with glaucoma alone. However, its sensitivity and specificity decreased to 0.375 and 1.0, respectively, in patients with concomitant RVO. In contrast, the RVO-GLA AI achieved an area under the curve (AUC) of 0.875, with a sensitivity of 0.87 and a specificity of 0.71. Across all the ophthalmologists, the average sensitivity was 0.63, and the specificity was 0.87. Specialists achieved a sensitivity of 0.80 and a specificity of 0.89, while residents had a sensitivity of 0.46 and a specificity of 0.85. Conclusions: An AI-based clinical decision support system specifically designed for glaucoma detection significantly improved diagnostic performance in eyes with combined RVO and glaucoma, achieving an accuracy comparable to that of ophthalmologists, even with a limited number of training cases.