Development of a cloud-based estimator for analysing the disc-fovea angle in fundus images using a stacking ensemble model.

PURPOSE: To develop a cloud-based software for analysing the disc-fovea angle (DFA) in fundus images using the stacking ensemble model and to evaluate the accuracy of the objective DFA assessment between automatic and manual analyses.

METHODS

A total of 682 fundus images (573 and 109 images of healthy individuals and patients with cyclotropia, respectively) were included in this retrospective study. For manual DFA analysis, two examiners analysed the DFA using web-based analysis software for the DFA, which was defined as the average value between the two examiners. For automatic DFA analysis, the stacking ensemble model on the cloud web integrated a single-shot multi-box detector to identify the fovea and centre of the optic nerve head, along with a single convolutional neural network combined with five machine learning algorithms.

RESULTS

The DFA did not significantly vary between the automatic (7.25° ± 4.93°) and manual analyses (7.27° ± 5.07°) (p=0.52). The time required to analyse the DFA in a single fundus image was significantly shorter with the automatic analysis (0.430±0.240 s) than with the manual analysis (15.794±1.558 s) (p<0.001).

CONCLUSION

These findings suggest that automatic analysis using the stacking ensemble model is useful for determining the DFA in clinical settings.