Abstract
Purpose:
To develop an automated progression analysis method of retinal vessel wall thickness using registered adaptive optics scanning laser ophthalmoscope (AOSLO) videos and corresponding pulse wave data.
Methods:
First of all, we capture AOSLO videos and pulse wave data in different examination dates for each subject. When a baseline video and a measurement position are selected, our method automatically calculate wall-to-lumen ratio (WLR) in the corresponding measurement position for each follow-up video based on the following procedure; (i) registration of each follow-up video to a reference frame of baseline video (ii) selecting frames with a predefined relative cardiac cycle value and with high contrast from each stabilized video (iii) producing smoothed image with preserved edge from the frames selected in (ii) by edge-preserving filtering (iv) edge detection of retinal vessel wall and calculating WLR for each follow-up data, followed by showing the progression graph. In this study, we recorded AOSLO videos using Canon prototype system with a high wavefront correction efficiency using dual liquid crystal phase modulator. The imaging light wavelength was 840 nm and the frame rate was 32 Hz. The scan area around the optic disc was 2.8 × 2.8° and was sampled at 400 × 400 pixels. The videos were recorded for 4 seconds in each examination. We investigated inter-frame registration precision by placing 15 blocks in each frame and calculate shift values of them between reference frame and other frame using 4 follow-up AOSLO videos for each healthy subject (3 subjects in total). We also investigated residual error between automatically measured WLR and manually measured one by an expert operator using 7 AOSLO videos for 6 healthy subjects.
Results:
The shift value for each block between the reference frame of baseline video and each selected frame of other video was 0.085 pixels on average. In consideration of inter-examination registration, we successfully specified corresponding measurement position between follow-up data by image warping and relative cardiac cycle selection. The mean error of WLR measured by the automated method was significantly lower than the one measured by non-expert operator (p<0.05).
Conclusions:
We developed an automated progression analysis method of retinal vessel wall thickness from AOSLO videos, which enabled easy and accurate assessment of progression.
Keywords: 549 image processing •
688 retina