Abstract
Purpose :
Arita et al. demonstrated a simple and highly effective diagnosis to classify dry eye disease (DED) type to normal (healthy), aqueous deficient dry eye (ADDE) and evaporative dry eye (EDE) types by focusing on the type-unique appearances of interfering fringe color (IFC) images (IOVS, 2016, 57, 3928). However, diagnostic biases by unskilled observers have been issues to be solved. In this study we built machine learning (ML) models that use artificial intelligence (AI) technologies to classify IFC images. And we examined if the ML models have abilities to support DED type diagnosis.
Methods :
IFC image taken by a tear interferometer Kowa DR-1α at 5 seconds after blink was used. We collected 46 images from each type (138 images in total). The ML models were trained by training image sets (randomly chosen 31 images from each type; 93 images in total) and DED type of test images (residual 15 images from each type; 45 images in total) were predicted by ML models. We prepared two ML models for the type prediction; 1) "bag of visual words" (BVW) method, a widely used standard way to classify images; 2) our original method that uses predefined 11 image features (sets of numerical values) specially designed for IFC image appearances including brightness of each color channel, color ratio, color saturation, local complexities of image, etc. F-scores and kappa coefficients of predicted types by ML models against those by skilled ophthalmologists were evaluated to know the efficacies of the ML models in DED type classification.
Results :
Our ML model showed effectively high F-scores (normal, 0.845 ± 0.067; ADDE, 0.981 ± 0.023; EDE, 0.815 ± 0.095) and inter-rater agreement (kappa coefficient = 0.820) against predetermined DED types by skilled ophthalmologists, even though BVW method showed very low F-scores (normal, 0.587 ± 0.146; ADDE, 0.669 ± 0.073; EDE, 0.549 ± 0.088) and agreement (kappa coefficient = 0.367).
Conclusions :
Our ML model showed almost perfect agreement in the classification of DED types. This result indicates that ML model would sufficiently support the diagnosis of DED types and that the model would reduce the diagnostic biases by observers. In the next step we will confirm this through the clinical evaluations using many IFC image datasets from DED patients.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.