Abstract
Purpose :
Adaptive Optics Scanning Light Ophthalmoscopy (AOSLO) provides exceptional resolution and contrast of photoreceptors in a clinical setting. However one barrier to wide spread clinical adoption of this imaging modality is the time required to acquire and process images. For example, imaging a 1x10 degree region takes tens of minutes, while processing the data to produce a montage can take several person days. We hypothesize that optical flow could be used to rapidly acquire and process large panoramic images by acquiring video frames while slewing the fixation target.
Methods :
ASLO videos (1.5 degree feild of view, 13 frames per second) were acquired while continuously moving the fixation target at a rate of 2 to 3 degrees per second to produce overlapping frames. The slew rate of the fixation target can be chosen to produce the desired number of frames for each retinal location for sufficient noise averaging. A custom python scripts was written to calculate optical flow using OpenCV library as well as a custom optical flow based on normalized cross correlation (NCC). Flow values were used to align and unwarp individual video frames and the average was calculated to produce a panoramic image.
Results :
Open source libraries for caluclating optical flow based on features performed poorly due to the lack of edges and corners present in AOSLO images. Dense optical flow methods showed promise, but were not optimized for the motion artifacts common in AOSLO images. A custom NCC based flow calculation of strips provided the best results and enabled rapid alignment and dewarping of 60 frames in 60 seconds to produce 1x10 degree FOV.
Conclusions :
One important barrior to clinical adoption of cellular scale retinal imaging with AOSLO is the time required to acquire and process images. Implementing compuational methods like optical flow to automate the stitching of video frames could speed up both acquisition and processing of large aread of retina in minutes making this important imaging modality more emmenable to clinical applications.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.