Abstract
Scanning laser ophthalmoscopy (SLO) recordings can be contaminated with artifacts caused by eye movements. Artifacts cause both distortion within an SLO frame and motion between frames of an SLO movie. These cause problems when analysing SLO movies, when averaging frames to decrease noise or tracking features over time. Many algorithms exist for spatially registering images; this study develops well characterised datasets to compare objectively the performance of two full frame registration algorithms.
SLO movies were generated by applying eye movement data, recorded using a pupil tracking system, to a static image recorded from an adaptive optics enhanced SLO (AOSLO). Eye movements were recorded from both a participant with nystagmus and a control at 200Hz, then interpolated to the pixel sampling rate of the AOSLO. Two algorithms were used to align frames in the resultant movies. The first algorithm relies on feature alignment using a pyramid search. The second performs image registration in the Fourier domain. Performance was measured by creating an average frame from the registered movie. Metrics included: Signal to noise ratio (SNR); Acutance, a measure of edge contrast, and Entropy, a measure of randomness in an image [adapted from Hunter et al. Characterizing image quality in ascanning laser ophthalmoscope with differing pinholes and induced scattered light. J Opt Soc Am A. 2007].
The addition of eye movements to the static image caused a reduction in acutance (p<<0.001). Both registration methods caused an improvement in all performance metrics (p<0.001) after addition of eye movements. Acutance was greater after feature based registration than Fourier registration (p<0.05).