Artificial intelligence (AI) can accurately identify diabetic retinopathy in resource-limited clinical settings, but certain ocular characteristics may significantly reduce its diagnostic sensitivity, according to a study published in the Journal of Medical Internet Research.

The findings suggest that although automated retinal screening could help address shortages of ophthalmologists, clinicians should remain alert to factors that may obscure retinal lesions and increase the likelihood of missed diagnoses, particularly in patients with early-stage disease.

Led by Ying Xue, MD, of the Affiliated Hospital of Nantong University in China, the cross-sectional study evaluated the performance of the EVisionAI automated screening system in detecting diabetic retinopathy and vision-threatening diabetic retinopathy among adults with type 2 diabetes in rural Funing County, China.

Evaluating screening accuracy

Between October 21 and November 12, 2024, researchers enrolled 1,847 adults with type 2 diabetes, of whom 1,748 completed screening. Following the exclusion of 104 eyes with images unsuitable for analysis, 3,392 eyes were evaluated.

Participants underwent comprehensive ophthalmic examinations, including visual acuity testing, intraocular pressure measurement, slit-lamp examination, axial length measurement, and fundus photography. Two ophthalmologists independently graded retinal images according to American Academy of Ophthalmology standards, providing a reference against which EVisionAI's performance was assessed.

Diabetic retinopathy was identified in 19.2% of participants, while 7.4% had vision-threatening disease. Overall, EVisionAI demonstrated 90.61% sensitivity and 98.99% specificity for detecting diabetic retinopathy. Performance was particularly strong for vision-threatening disease, with sensitivity reaching 97.45% and specificity reaching 98.95%. However, sensitivity fell to 78.46% for early-stage diabetic retinopathy, highlighting limitations in identifying subtle retinal abnormalities.

When ocular characteristics interfere

The researchers identified three important factors affecting diagnostic performance: reduced image clarity, fundus tessellation, and vitreous degeneration.

Severe image clarity impairment, particularly around the optic disc, reduced overall diagnostic sensitivity to 80.95%. Marked fundus tessellation, characterized by increased visibility of underlying choroidal vessels, also affected detection, reducing sensitivity to 82.86%.

These effects were especially pronounced in early-stage disease. Among eyes with severe fundus tessellation, EVisionAI detected just 60.7% of early-stage diabetic retinopathy cases, compared with 79.8% in eyes with less pronounced tessellation.

The researchers suggested that visible choroidal vessels and media opacity may obscure or resemble subtle lesions, including microaneurysms and small retinal hemorrhages. Severe vitreous degeneration presented another challenge, potentially obscuring hemorrhages or causing vitreous abnormalities to be mistaken for retinal exudates. Sensitivity for diabetic retinopathy was higher in eyes with dilated pupils than in those without dilation (91.37% vs 87.41%), although specificity differed little between the groups, and dilation was not associated with improved early-stage detection.

Implications for clinical practice

The findings support AI-assisted screening as a potentially valuable approach for expanding diabetic eye care in regions with limited specialist availability, which the authors said could help alleviate pressure on diabetic retinopathy screening programs in China.

However, the authors also emphasized the importance of additional assessment by an ophthalmologist when severe vitreous abnormalities, media opacity, or marked fundus tessellation are present, particularly when automated results appear reassuring despite clinical concerns.

The study was limited by its focus on a single AI system and a predominantly middle-aged and older rural Chinese population with lower educational attainment, potentially restricting generalizability.

The researchers concluded that screening protocols should account for ocular characteristics that compromise automated detection. Further validation across different AI platforms and more diverse patient populations will be essential to improving diagnostic reliability and ensuring that patients requiring specialist assessment are not overlooked.

The authors declared no conflicts of interest.