Fig 1.
Ground truth data and generation of distorted SLO imagery.
a) Cone photoreceptor mosaic 1.25° temporal to the fovea, generated from the average of 1000 frames acquired at 750 nm using flood-based adaptive optics. White box indicates the extent of the 400x400 pixel (192x192 μm) simulated SLO frame. b) Eye movement data in pixels (1 pixel = 0.48 μm) from the same sequence that was used to generate a. Eye motion was determined from cross-correlation of each of 1000 frames, measured at 200 fps, with the first frame in the sequence. c) One of the SLO frames simulated at 20 fps, generated by combining ground truth image and eye movement data (corresponding to time point indicated by magenta lines). Distortion is evident towards the bottom of the frame, which corresponds to a spike in eye movement. S1 Video shows all simulated frames for this sequence, after full-field registration of each frame to the first frame.
Fig 2.
Depiction of two strategies to place rectangular strips in the correct position to minimize intra-frame distortion.
Top: a simple procedure in which each strip in the reference frame is compared to the corresponding strip, relative to the beginning of the frame, in each other frame of the sequence. Here no special effort is made to locate the corresponding anatomy; when displacement becomes too great (frame #3), the displacement cannot be recovered or can be recovered only poorly, leading to a divergence between the methods for larger displacements. Bottom: As for top, but now a robust registration procedure is used which ensures that the corresponding anatomy is correctly located, if present in the imaged field.
Fig 3.
Image similarity to the ground truth image for 100 frames recovered from a sequence.
Black: Ideal case in which the ground truth image is supplied as the reference for recovery, showing the upper limit of performance given the finite strip dimensions. Blue: Selection of a key frame, assuming perfect registration to that frame is achieved. Green: Simplified de-warp algorithm, in which eye movements made during the reference frame are determined and compensated by comparing strips acquired at the same time relative to the beginning of their frame (as described in Fig 2, top). Red: A combination of the above two methods. Grey line: Example frame compared for each method in S2 Video. Frames were re-ordered by increasing image similarity to aid visualization.
Fig 4.
Performance as a function of increasing amplitude of fixational eye movements.
Interquartile range (central 50% of data), for correlation of all recovered frames to the ground truth image, is plotted as a function of amplitude of eye movement.