aceArtificial intelligence for diagnosis of keratoconus using Scheimpflug based corneal tomography
Aim To develop and evaluate the diagnostic accuracy of deep learning (DL) models in differentiating keratoconus (KC) from normal eyes with regular astigmatism. Methods A comparative cross-sectional study was conducted at the Cornea and Diagnostic Department of Al-Shifa Trust Eye Hospital, Pakistan. Galilei dual Scheimpflug-based corneal topography was performed to obtain four corneal maps: anterior axial curvature, posterior axial curvature, corneal thickness, and posterior elevation. Four convolutional neural n…
"Doctors Hospital from the Southwest 1" by Sixflashphoto is licensed under CC BY-SA 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/4.0/.
By July 2026, a cross-sectional study at Al-Shifa Trust Eye Hospital in Pakistan developed four CNN models on 5602 Scheimpflug-derived corneal maps from 1411 eyes to distinguish keratoconus from normal eyes, reporting internal accuracies of 98.1% to 99.2% and AUCs up to 1.00, with external validation on 85 participants confirming 97.1% to 98.3% accuracy.
High-accuracy automated detection matters because early keratoconus identification guides timely crosslinking and refractive management, yet the reported performance is confined to a 10-40 year age range, a single institution, and one imaging platform, leaving open questions about generalizability, prospective clinical workflow integration, and impact on patient outcomes.
- Study was comparative cross-sectional at Cornea and Diagnostic Department of Al-Shifa Trust Eye Hospital, Pakistan using Galilei dual Scheimpflug topography.
- Dataset comprised 5602 corneal maps from 1411 eyes of 827 participants aged 10 to 40y, including 790 KC and 621 normal eyes.
- Four models were trained on anterior axial curvature, posterior axial curvature, corneal thickness, and posterior elevation maps.
- External validation used independent dataset of 85 participants, 150 eyes, 1050 maps.
Convolutional neural networks trained on four Scheimpflug-based corneal maps differentiated keratoconus from normal astigmatic eyes with up to 99.2% accuracy and AUC 1.00, with external validation retaining 97-98% accuracy.
The rundown
Researchers extracted four map types per eye from Galilei dual Scheimpflug tomography and trained DenseNet-121, ResNet-50, Inception-V3, and EfficientNet-B0 to classify KC versus normal regular astigmatism, evaluating AUC, accuracy, sensitivity, and specificity.
Internal testing showed DenseNet-121 at 99.2% accuracy and ResNet-50 at 99.0% with both reaching AUC 1.00, while external validation on 150 eyes maintained high performance, supporting potential clinical implementation for optimized KC management.
Sources
- Peer-reviewedInternational Journal of Ophthalmology2026-07-18
How should this claim be treated?
The debate