Abstract
Purpose :
Non-confocal split detector based adaptive optics scanning light ophthalmoscope imaging has been shown to reveal the cone photoreceptor inner segment even when the corresponding outer segments are not visible on confocal AOSLO. However, quantitative analysis of split detector AOSLO images currently requires the time-consuming manual marking of cones. In this work, we present the first fully-automated method for the detection of cones in SD-AOSLO images of healthy eyes.
Methods :
The algorithm works in two distinct steps. In the first step, non-pertinent information is removed by applying an adaptive band pass filter with parameters determined by estimating the location of Yellott’s ring in Fourier space. In the second step, a priori information about split detector images is used to aid in the local detection of cones, namely that cones appear as a high intensity spot next to a low intensity spot. Thus, local minima and maxima are paired together in order to detect cones. The parameters for our algorithm were trained on 32 images from 4 healthy eyes, and validated against the current gold standard of manual segmentation on an additional 80 images from 10 subjects containing 10500 cones. One-to-one pairs between the automatic and manual markings were found in order to calculate measures of the sensitivity and precision. A second manual grading of the validation set was also compared to the first to get measures of inter observer variability.
Results :
Figure 1 shows qualitative examples of the automatic segmentation compared to expert manual marking. Compared to the first manual grader, the automated method had an average sensitivity of .960 and precision of .947. Compared to the first manual grader, the second manual marking had a sensitivity of .956 and precision of .936. The average computation time was 0.05 seconds per image.
Conclusions :
Our fully automatic method for detecting cones in split detector images was congruent with manual segmentation. Differences between automatic and manual were shown to be comparable to differences between two manual graders.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.