Abstract
Purpose :
Adaptive-optics scanning laser ophthalmoscopy (AOSLO) records line-by-line video sequences that are subject to motion artifacts. Strip alignment cross-correlation has been the procedure of choice to correct for eye motion. The procedure requires human interaction in finding a reference frame. This procedure is time consuming and labor intensive. Typical estimates are that we need about 8-10 hours of processing time for every hour of subject imaging time.
Methods :
AOSLO video sequences were recorded with our custom-built instrument using the design by Dubra. Images were exported to a workstation with nVidia multicore graphics card running Matlab and Python. Conventional manual alignment was performed using custom-built programs that select registration parameter and subsequently run the cross-correlation algorithms using the CUDA language on the nVidia graphics card. Automatic alignment was programmed in Matlab using several steps. An interative process was used to find the best reference frame that gave the best correlation coefficient with the entire image set. We have used the Heidelberg High Magnification Module (HMM) to screen patients for clear media.
Results :
Manual registration takes about 10 min per image set. The automatic reference frame process takes about 4-5 minutes on average. The entire strip-alignment process takes about 10 minutes per image set and is fully automated to process a folder at a time. This allows off-line overnight processing of the entire data set without user interaction. Using the HMM helped us determine probability of success in achieving adequate image quality.
Conclusions :
AOSLO is limited in its clinical application by the small field-of-view, long-imaging time and long post-acquisition processing time. This severely limited the ability of clinicians to use AOSLO imaging on a routine basis. The small field-of-view requires that a large number of images have to acquired to yield a reasonable mosaic of overlapping individual images. In addition, light safely limits the quality of a single frame making it necessary to acquire a series of sequential frames at the same location. Continued advances in image processing and hardware motion correction are required to transform the operation of AOSLO imaging into a routine clinical device.
This is a 2020 ARVO Annual Meeting abstract.