A novel objective method to detect the foveal center point in the rtx1^TM device using artificial intelligence

Abstract

Purpose : rtx1^TM is a valuable imaging device for evaluating retinal microstructures. While precise localization is essential, the lack of a gaze tracker and registration system makes device utilization challenging in clinical practice. Currently, a subjective method is employed for the device to detect the central foveal point. Herein, we devised a novel objective method to determine the foveal center point of the eyes using artificial intelligence (AI).

Methods : Seventy-one eyes (38 subjects) were enrolled in our study. Images of five regions of interest (ROI) (4° x 4°) with the central points of (0, 0), (0, -2), (0, +2), (2N, 0), and (2T, 0) were captured and montaged using adaptive optics retinal camera, rtx1^TM. Spectral-domain optical coherence tomography (SD-OCT) images of all eyes were also captured and considered as a reference to determine the precise location of the foveal center. Two clinicians manually delineated the borders of the fovea based on the blurriness area using both a. tangential (Tg) and b. best-fit sphere (BFS) methods (Figure 1). We also superimposed the montage image of rtx1^TM on the images of en face OCT manually and marked the OCT-based foveal (OBF) center. We ran the deep learning regression-based model to predict the center points on rtx1^TM images based on manually overlayed OCT centers. Ultimately, these predicted AI centers are compared with the OBF, Tg, BFS, and device-based (DB) centers.

Results : The mean age of the participants was 30.9±6.2, and 35% were females. Median distances between OBF points with DB, Tg, BFS, and AI center points were 110.345 (22.28-210.27), 102.44 (69.85-170.3), 106.842 (74-145.82), and 96.2(48.99-134.44) μm, respectively (Figure 2). There was a significant difference between AI and DB center distance from the OB point (p=0.0017). There was no significant difference between the AI center with Tg and BFS center distance from the OBF point (p=0.106,0.06). There were significant differences between the DB with Tg and BFS centers’ distance from the OBF point (p=2.22x10^-10, 3.86x10^-10, respectively).

Conclusions : A novel AI-assisted localization is a helpful method to objectively determine the foveal center quite precisely in AO images which may cover the weakness of the currently used method and may also be beneficial in follow-up-image evaluation.<gwmw style="display:none;"></gwmw>

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