Abstract
Purpose :
Robust and real-time retinal tracking is a crucial component of an optical coherence tomography (OCT) acquisition system. We propose a real-time tracking algorithm, based on deep-learning and computer vision techniques, that uses an anatomical feature such as the optic nerve head (ONH) as a reference point to improve the tracking performance.
Methods :
The tracking parameters (xy-translation and rotation) are calculated by registration between a reference image (RI) and a moving image (MI). Our method requires the ONH location and a set of RI landmarks extracted from the feature rich areas. The ONH in a RI is detected using a U-Net architecture. The ONH in a MI is detected by template-matching using ONH template extracted from the RI. Each reference landmark template and its relative distance to the ONH are used to search for corresponding moving landmark with the same distance from the ONH in MI (Fig 1). A subset of landmark correspondences with high confidence is used to compute the tracking parameters. IR images (11.52x9.36 mm2 with a pixel size of 15 µm/pixel) using a CLARUSTM 500 (ZEISS, Dublin, CA) at 50 Hz frame rate were collected, using normal and small pupil acquisition modes with induced eye motion. Each eye was scanned using 3 different motion levels; good fixation, systematic, and random eye movement. The registered images were displayed in a single image to visualize the registration (Fig 1). The mean distance error between the registered moving and reference landmarks was calculated as the registration error. The statistics of registration error and eye motion were reported for each acquisition mode and each motion level.
Results :
Around 500 images each were collected from 15 eyes. Fig 2 shows the statistics for the registration error and eye motion for different acquisition modes and motion level using all eyes. The mean and standard deviation of registration error for normal and small pupil acquisition modes are similar which indicates that the tracking algorithm has a similar performance for both modes. Reported registration errors are important information which help to design an OCT scan pattern. The tracking time for a single image was measured 13 ms on average using Intel i7 CPU, 2.6GHz, 32GB RAM.
Conclusions :
We developed a real-time retinal tracking method using IR images. We demonstrated relatively good tracking performance which is an important part of an OCT image acquisition system.
This is a 2020 Imaging in the Eye Conference abstract.