Purpose:
In vivo retinal imaging, such as with optical coherence tomography (OCT), is often realized by mechanical scanning. While sequential acquisition has many advantages, retinal motion during the acquisition can distort the image even under normal fixation. For high-resolution retinal imaging, as for example OCT with adaptive optics (AO-OCT), the magnitude of image distortion can be many times larger than the cellular structures of interest. Here, we propose a new registration algorithm for stabilizing motion-corrupted videos of cone photoreceptors acquired with scan-based AO-OCT.
Methods:
We acquired 12 AO-OCT[1] volume movies of 0.50x0.50 retinal patches obtained on four subjects. En face frames of the cones were extracted and then registered as follows: (1) manual selection of 10-20 landmark cones, (2) partitioning each frame into narrow strips, (3) registration of each strip using an anisotropic affine transformation (scaling in slow scan direction and shearing in fast scan direction), and (4) reassembling the registered strips. Effectiveness of the registration method was quantified by (1) measuring the lateral motion of cones before and after registration and (2) comparing to a conventional registration method based on thin-plate-spline (TPS).
Results:
The figure shows the average of 16 en face frames with and without registration (red crosses denote landmark cones). Without registration, cone misregistration between frames severely degrades the average image. TPS registration noticeably improves cone alignment between frames, but the proposed affine method is more effective. Registration error (root-mean-square distance from reference cone locations) of five randomly selected cones is 18.0±6.4 µm (no registration), 6.6±1.3 µm (TPS), and 2.6±0.8 µm (proposed).
Conclusions:
Proposed affine method is effective at registering cone photoreceptors to less than a cone width.[1] B. Cense et al., Opt. Express 17, 4095-4111 (2009).
Keywords: image processing • imaging/image analysis: non-clinical