Association for Research in Vision and Ophthalmology

Association for Research in Vision and Ophthalmology

Source:

Liu H, et al. Deep learning for detection of keratoconus and prediction of crosslinking efficacy. Presented at: Association for Research in Vision and Ophthalmology annual meeting; May 1-7, 2021 (virtual meeting).

Disclosures: Liu reports no relevant financial disclosures.
May 03, 2021
2 min read
Save

Deep learning model shows efficacy in detection, staging of keratoconus

Source:

Liu H, et al. Deep learning for detection of keratoconus and prediction of crosslinking efficacy. Presented at: Association for Research in Vision and Ophthalmology annual meeting; May 1-7, 2021 (virtual meeting).

Disclosures: Liu reports no relevant financial disclosures.
You've successfully added to your alerts. You will receive an email when new content is published.

Click Here to Manage Email Alerts

We were unable to process your request. Please try again later. If you continue to have this issue please contact customerservice@slackinc.com.

A deep learning model using topography scans showed the ability to effectively differentiate keratoconus vs. normal corneas with an accuracy of 99.3% and to stage the disease with an accuracy of 88%.

“Unfortunately, the program was unable to learn any useful parameter that would help predicting the likelihood of success vs. disease progression following corneal cross-linking, which may partially be attributed to the reduced sample size given the rarity in disease progression,” Henry Liu, MD, said at the virtual Association for Research in Vision and Ophthalmology meeting.

Deep learning with convolutional neural networks (CNN) has been used extensively in ophthalmology in recent years, mainly in the field of retina and glaucoma. So far, only one study has been conducted to diagnose keratoconus, and this was based on OCT images rather than topographies, which are by far the most widely used method, Liu said.

The study he presented was conducted at the University of Ottawa, Canada, and was aimed at implementing and testing a trained CNN deep learning algorithm using preoperative Pentacam (Oculus) topography scans in order to differentiate between keratoconus and normal corneas, staging the disease and predicting whether a patient will likely benefit from cross-linking treatment.

A total of 2,450 Pentacam scans were used, approximately half of which were from patients with a diagnosis of keratoconus and half from healthy controls undergoing refractive surgery. A CNN was implemented and trained using 80% of the data set for training and 20% for validation.

“Our deep learning algorithm was able to discriminate keratoconus from normal corneas with an accuracy of 99.3%. With the exception of corneal thickness alone, most of the Pentacam parameters individually (anterior curvature, anterior and posterior corneal elevation) were able to predict whether the cornea was keratoconic or not,” Liu said.

Staging based on the Amsler-Krumeich scale achieved a validation accuracy of 73.5%, which increased to 88% by incorporating manifest refraction data.

The final stage of the study was aimed at assessing the algorithm’s ability to predict the patient response to cross-linking.

“Progression of keratoconus, if detected early, can be effectively halted with CXL. However, in spite of treatment, 10% of patients continue to progress, eventually requiring corneal transplantation. It would be beneficial to have an automated method to predict the success of cross-linking,” Liu said.

The presence of only 72 cases of progression in the data set limited the ability of the algorithm to learn useful parameters for categorizing the scans as stabilized vs. progressing, yielding a final validation accuracy of only 53.6%.

“Even though it was not able to predict disease progression, our deep learning model showed that the applications of CNN are encouraging, can be utilized to solve practical clinical problems and serve as an adjunct to clinical decision-making in the future,” Liu said.