Abstract
Purpose :
To develop and validate a machine learning algorithm for accurate estimation of the optic disc and fovea center position in infra-red SLO fundus images including cases outside of the field of view or apparent occlusions of the landmarks.
Methods :
To detect the coordinates of obstructed or out-of-view optic disc or fovea, we reformulated the task as a regression problem where a machine learning algorithm learns to predict an anatomically plausible coordinate system centered on the fovea. The prediction was obtained by training a neural network with EfficientNet backbone. To this end, a dataset of 1226 grayscale OCT localizer images (Heidelberg Spectralis, average field of view: 8.75x8.75mm) affected with either Retinal Vein Occlusion related Macular Edema, Diabetic Macular Edema or Age-Related Macular Degeneration with choroidal neovascularization, was annotated and randomly split into a train (1106) and a test set (120). The test set included 46 images for which at least 50% of the optic disc was outside field-of-view.
An outlier-robust estimation (RANSAC) was used to determine the final fovea and optic disc location in an anatomical coordinate system. The detection was evaluated by computing the average distance between manual annotation and predicted location.
Results :
The presented methodology was able to estimate the location of the optic disc with an average error of 0.07 mm and the fovea with an average error of 0.2 mm, independent of the diseases (Fig. 1).
The method showed good performance also in cases where optic disc or fovea were only partially visible or fully-occluded due to existing lesion or artificial image cropping (Fig. 2 bottom). The goodness of fit was a good surrogate of the detection error (Fig. 2 top).
Conclusions :
The automated detection algorithm could identify fovea and optic disc locations with a high level of accuracy. The addition of RANSAC increased the robustness of the model also in the presence of occlusions and enabled quantification of the localization accuracy. This method allows the registration of longitudinal scans, even if the optic disc is not or only partially visible, as it is often the case on fovea-centered OCT scans.
This is a 2021 ARVO Annual Meeting abstract.