Deep Learning for Detection of Keratoconus and Prediction of Crosslinking Efficacy

Abstract

Purpose : Deep learning with convolutional neural networks (CNN), a method of supervised machine learning that uses multiple layers to progressively extract higher level features, is becoming increasingly popular in recent years in the field of ophthalmology. The goal of this study is to implement and test a trained deep learning algorithm using pre-operative corneal topography scans in order to 1) classify between keratoconus and normal corneas 2) stage the disease according to Amsler-Krumeich scale and 3) predict whether a patient will likely to benefit from crosslinking treatment.

Methods : Scans of baseline visits for all patients that had a diagnosis of keratoconus (case group) and those who have been assessed for refractive surgery (control group) between January 2007 and June 2019 were analyzed. A CNN was implemented and trained with our custom dataset of corneal topographies (with 80% and 20% validation). In total, 2410 scans were included (1163 keratoconus and 1247 controls) for stage 1, 985 keratoconus scans were classified according to the Amsler-Krumeich scale for stage 2, and 138 keratoconus scans (69 progressed and 69 stabilized) were labeled for stage 3.

Results : For stage 1, the deep learning model with data including all four pentacam parameters (anterior and posterior corneal elevations, anterior curvature and pachymetry maps) showed a validation accuracy of 0.995 in discriminating between keratoconic and normal corneas. For individual map analysis, corneal pachymetry showed the lowest validation accuracy of 0.771, whereas the other parameters anterior corneal elevation, posterior corneal elevation and anterior curvature maps showed accuracies of 0.987, 0.984 and 0.978 respectively. For stage 2, the algorithm was able to stage the disease with a validation accuracy of 0.735 (excluding manifest refraction as an input parameter) and 0.878 (including manifest refraction as an input parameter). For stage 3, the accuracy in correctly predicting the progression of keratoconus was 0.536.

Conclusions : Deep learning using topography scans obtained from the pentacam effectively differentiates keratoconus versus normal corneas as well as stages the disease according to established criteria. The algorithm was not able to predict the likelihood of disease progression based solely on the pre-operative topography, which may partially be attributed to the reduced sample size given the rarity in disease progression.

This is a 2021 ARVO Annual Meeting abstract.