Keratoconus Disease Prediction by Utilizing Feature-Based Recurrent Neural Network

: Keratoconus is a noninflammatory disorder marked by gradual corneal thinning, distortion, and scarring. Vision is significantly distorted in advanced case, so an accurate diagnosis in early stages has a great importance and avoid complications after the refractive surgery. In this project, a novel approach for detecting Keratoconus from clinical images was presented. In this regard, 900 images of Cornea were used and seven morphological features consist of area, majoraxislength, minoraxislength, convexarea, perimeter, eccentricity and extent are defined. For reducing the high dimensionality datasets without deteriorate the information significantly, principal component analysis (PCA) as a powerful tool was used and the contribution of different PCs are determined. In this regard, Box plot, Covariance matrix, Pair plot, Scree Plot and Pareto plot were used for realizing the relation between different features. Improved recurrent neural network (RNN) with Grey Wolf optimization method was used for classification. Based on the obtained results, the average prediction error of the visual characteristics of a patient with keratoconus six and twelve months after the Kraring ring implantation using RNN are 9.82% and 9.29%, respectively. The average error of estimating characteristics of predicting the visual characteristics of a patient with keratoconus six and twelve months after myoring ring implantation are 11.46% and 7.47% respectively.