Automatic Estimation of Objective Cyclodeviation in Fundus Image Using Machine Learning

Abstract

Calculating the angle of objective cyclodeviation has been based on the difference between the fovea and center of the optic nerve head (ONH) in fundus images. An examiner defines the location of the fovea and center of ONH manually; thus, there is an influence of bias, and the analysis time per fundus image is long. Therefore, in this study, we developed a machine learning-based model for automatic cyclodeviation estimation using fundus images. Our model is divided into two steps: detecting macular and ONH regions using the single shot multibox detector (SSD) and estimating cyclodeviaiton using gradient boosting on decision trees of CatBoost. The SSD detected the macular and ONH with a 75% average precision of 99.7% and 99.4%. Then, CatBoost estimated the objective manual cyclodeviation with a mean absolute error of 0.804. These findings suggest that the machine learning-based algorithm estimate cyclodeviation with the same accuracy as a human from fundus images.