nnU-Net Pipeline Offers Objective Optic Disc Tilt Classification in Fundus Photographs

An automated AI pipeline, externally validated on 2,448 fundus photos, classified optic disc tilt with a clinical acceptance rate >98%. The tool detected tilt in 7.5% of images overall and 9.7% of normal eyes, but performance dropped for images with edema, with acceptance as low…

• This retrospective study developed an nnU-Net model for disc segmentation, validated on the SMDG dataset (n=3,103) and the external SMC dataset (n=2,448 photos from 1,370 patients), with two reviewers assessing clinical acceptance.
• Tilt classification used an elliptical ratio threshold of ≥1.3; tilt prevalence varied from 3.9% in glaucoma to 14.2% in eyes with edema.
• Primary limitation was reduced segmentation accuracy in the presence of disc edema and peripapillary atrophy, where acceptance rates fell to 81.6–93.9%.

RadPigeon summaries are original and for information only. They are not clinical advice.