Abstract: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 network models were developed and trained on corneal maps to classify eyes as KC and normal. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. RESULTS: A total of 5602 corneal maps were extracted from 1411 eyes (790 KC and 621 normal) of 827 participants, including KC (472) and normal (355) groups, aged 10 to 40y. The DL models achieved the highest accuracy with DenseNet-121 (99.2%), ResNet-50 (99.0%), Inception-V3 (98.6%), and EfficientNet-B0 (98.1%). DenseNet-121 and ResNet-50 achieved an AUC of 1.00. External validation on an independent dataset of 85 participants (150 eyes with 1050 extracted corneal maps) confirmed excellent accuracies for EfficientNet-B0 (98.1%), DenseNet-121 (98.3%), and ResNet-50 (97.1%). CONCLUSION: All DL models demonstrate excellent diagnostic accuracy for KC detection, highlighting the potential for clinical implementation and optimized KC management with greater precision.
