Intelligent Classifiers Of Keratoconus In Small-Diameter Corneas Based On Deep Learning Algorithm

To establish deep learning classifiers of keratoconus for small-diameter corneas by data mining and analysis of Pentacam tomography data and to evaluate the accuracy of classifiers in discriminating healthy corneas from the suspecious corneas.

Aier School of Ophthalmology, Central South University, Changsha, Hunan province, China.

Patients with corneal diameter ≤ 11.1 mm measured by Pentacam HR Scheimpflug camera were labeled as normal corneas, suspect keratoconus, and keratoconus groups based on clinical manifestation and the Belin/Ambrósio enhanced ectasia display (BAD). The classifiers were built by ResNet (Residual Network), ViT (Vision Transformer), and BotNet (Bottleneck Transformer), respectively. Receiver operating characteristic (ROC) was used to evaluate the performance of classifiers. Through multi-center clinical studies, 585 patients from seven other hospitals were included for model validation to compare the diagnostic ability of the best classifier and of the BAD system for keratoconus.

The accuracy of ResNet, ViT, and BotNet for diagnosis of normal cornea and suspect keratoconus was 85.57%, 86.11%, and 86.54% respectively, and the area under ROC of the test set was 0.823, 0.830 and 0.842 respectively. The accuracy of classifiers for diagnosis of suspect keratoconus and keratoconus was 97.22%, 95.83%, and 98.61%, respectively, and the area under ROC of the test was 0.951, 0.939, and 0.988 respectively. The false positive rate of BAD system was 46.05%, and the false positive rate of BotNet was 8.40%.

For corneas ≤11.1 mm in diameter, the BotNet has the best accuracy rate for classifying keratoconus, which can reduce the false positive rate of BAD system and provide real and effective guidance for early diagnosis.