Abstract
Purpose :
Macular Thickness Analysis (MTA) is a widely used tool for diagnosing and monitoring patients with ocular pathologies. The robustness of MTA is directly connected to the location of the fovea, which is used to place the ETDRS grid on the macular thickness (MT) map. The fovea localization based on OCT data becomes problematic when the OCT data is generated by a low-cost OCT system due to low contrast in the data. We developed a deep learning-based algorithm to localize the fovea using MT maps for low-cost OCT.
Methods :
Scans from one or both eyes of 562 subjects (with one or more scans) were used for training and validation. In each case, the resulting OCT volumes over 6×6 mm using PLEX® Elite 9000 (ZEISS, Dublin, CA) and CIRRUS™ HD-OCT 5000 with/without AngioPlex® OCT Angiography (ZEISS, Dublin, CA) were segmented to delineate the inner limiting membrane (ILM) and the retinal pigment epithelium (RPE). The prototype segmentation was used to generate MT maps with 512x512 pixels over an area of 5.78×5.78 mm. The principal curvature maps (min and max) were calculated using the MT map (Figure 1). One grader manually selected the center of fovea used as the ground truth. This resulted in 1020 samples, from which 816 are used for training and 204 for the first test set (i.e., high-quality set). Additionally, 492 samples (extracted from multiple scans, one or both eyes of 82 subjects) from a prototype low-cost device formed a second test set. No samples from this low-cost device are used for training.
The MT map and corresponding curvatures maps were input to a MobileNetV2, trained to estimate a single point. Various geometric and photometric augmentations were used during the training to increase the model's generalization.
Results :
The difference between ground truth and estimation was 65±71 microns for the first test set of 204 samples from the same devices as training data. Error on test data from the low-cost device was 100±138 microns (see Figure 1-d). Figure 2-a shows the error distribution on low-cost OCT data and other mentioned devices. Figure 2-b shows the model's accuracy for various error thresholds. Processing time was 40 ms on an Intel Xeon CPU and 15 ms using a GTX 1080Ti GPU.
Conclusions :
The result showed the effectiveness and robustness of the method to locate the fovea from MT maps obtained using low-cost OCT data.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.