Readers interpret the fundus images with AI-generated outputs available.

Readers interpret fundus images without access to the AI system.

The intervention consists of an AI-based fundus image interpretation software that provides automated outputs for 12 retinal and optic nerve findings (e.g., hemorrhage, exudates, drusen, optic disc change). The system does not generate a direct disease diagnosis. Instead, the AI displays the presence or absence of 12 predefined findings along with their lesion locations. Readers may use this finding-level information as decision-support when determining the presence of the four target diseases (diabetic retinopathy, age-related macular degeneration, retinal vein occlusion, and glaucoma).

Sensitivity of reader diagnoses for each of the four target diseases (DR, AMD, RVO, glaucoma) and for any fundus abnormality will be assessed with and without AI assistance, using the image-level reference standard as the comparator, through two reading sessions in which all 10 readers review all cases-randomised for each reader-with a washout period implemented to mitigate recall bias.

Specificity of reader diagnoses for each of the four target diseases (DR, AMD, RVO, glaucoma) and for any fundus abnormality will be assessed with and without AI assistance, using the image-level reference standard as the comparator, through two reading sessions in which all 10 readers review all cases-randomised for each reader-with a washout period implemented to mitigate recall bias.

Reading time per image will be measured during both unassisted and AI-assisted interpretation sessions. For each case, the total time from the moment the image is displayed to the moment the reader submits the final disease classification will be recorded automatically by the reading platform. Mean reading time per image will be calculated for each reader and compared between the two conditions to evaluate whether AI assistance reduces interpretation time.