Development of a cloud-based approach involving machine learning to estimate cyclodeviation

Abstract

Purpose : The calculation of objective cyclodeviation based on the difference between the fovea and center of the optic nerve head (ONH) in fundus images helps to diagnose cyclotropia. However, the analysis time per fundus image in the conventional approach is quite long because the examiner visually checks the fundus image. Machine learning algorithm has recently achieved excellent image processing results. In computer science, it is essential to understand programming language and set up a development environment with the necessary libraries; however, many medical workers have difficulty with computer science. Therefore, this study aimed to (1) develop an algorithm for the automatic estimation of objective cyclodeviation in a cloud-based application to make it accessible to everyone over the internet and (2) evaluate the accuracy of objective cyclodeviation between the automatic and conventional manual approaches.

Methods : The 573 fundus images were randomly divided into training (n = 458) and test (n = 115) data sets. Two examiners double-checked cyclodeviation in all fundus images using Cyclocheck^®, a web-based application to measure objective cyclodeviation manually (manual cyclodeviation). Google Colaboratory was used as the cloud-based application. The deep learning-based object detection algorithm used a single-shot multibox detector to detect the macular and ONH regions. Temporary cyclodeviation was defined as the difference between the center of the macular and ONH regions. Then, objective automatic cyclodeviation (automatic cyclodeviation) was estimated by gradient boosting on decision trees with CatBoost.

Results : There were no differences in the fixed (P = 0.89) and proportional (R² = 0.034, P = 0.096) biases between automatic (6.40 ± 3.35 degrees) and manual cyclodeviation (6.36 ± 3.58 degrees). The analysis time was significantly faster for automatic cyclodeviation (0.802 ± 0.268 s/image) than for manual cyclodeviation (15.798 ± 1.668 s/image) (P < 0.001).

Conclusions : There was no systematic error for objective automatic cyclodeviation when compared with conventional manual cyclodeviation. This finding suggests that the cloud-based application with the developed algorithm could be used by anyone with access to the internet to more quickly determine objective cyclodeviation automatically in fundus photographs when compared with the conventional manual approach.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.