Abstract: ObjectiveTo investigate the efficiency of artificial intelligence （AI）-assisted non-dilated fundus photography in diabetic retinopathy （DR） screening in a real-world clinical setting and evaluate its diagnostic consistency with ophthalmologists’ assessments. MethodsIn this prospective observational study, 14, 305 type 2 diabetes mellitus （T2DM） patients who underwent non-dilated fundus examination at the Metabolic Disease Management Center （MMC） of Tianjin Fourth Central Hospital between October 2018 and December 2024 were enrolled. The AI system （VoxelCloud） was used to analyze images captured during non-dilated fundus photography. The weighted Kappa test was employed to assess the agreement between the AI system and expert ophthalmologists. Screening failure rates and causes were also analyzed. ResultsThe overall DR prevalence was 21.4% （3, 056/14, 305）, with a DR positivity rate of 17.2% among patients with T2DM duration of <1 year. The AI system demonstrated substantial agreement with ophthalmologists Kappa = 0.817（95%CI 0.797-0.838）, P < 0.001. For moderate-to-severe DR, the AI system achieved a sensitivity of 97.3% and specificity of 95.9%. The screening failure rate was 3.7% （115/3, 085）, primarily due to small pupil size （70.4%） and media opaque caused by conditions such as cataracts （24.3%）. ConclusionImplementing non-dilated fundus photography in endocrine clinics facilitates early DR screening. AI-assisted non-dilated screening demonstrates high efficacy in real-world clinical practice.