Purchase this article with an account.
Keiko Yonezawa, Hiraku Sonobe, Akihito Uji, Sotaro Ooto, Masanori Hangai, Nagahisa Yoshimura; Improved Sensitivity of Automated Cone Identification in Images of Adaptive Optics Scanning Laser Ophthalmoscope (AOSLO). Invest. Ophthalmol. Vis. Sci. 2013;54(15):5551.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To provide an automated, highly sensitive cone detection algorithm even in the region where some cones are undetectable by the conventional local maximum detection algorithm previously proposed by Li and Roorda.
In retinal images taken by our AOSLO, some cones were not detected by the conventional algorithm because those cones did not have local max intensity profile by the effect of high intensity neighboring cones. The undetected cones were precisely picked up with our newly developed technique which combined a filter to analyze local convex profile with the conventional local max filter. Retinal images of healthy subjects were acquired with our AOSLO, all with resolution of 400 x 400 pixels and 340μm x 340μm in size. Total six images at eccentricities of 0.5mm (3 images) and 1.0mm (3 images) were selected for analysis. Cones in the images were first detected by the conventional algorithm, and undetected cones were manually labeled by two engineers. A pair of cones labeled manually by the different engineer was considered the same when they were located within 1.2μm for 0.5mm and 1.7μm for 1.0mm of each other. Comparisons were made between the result of detection using our algorithm and manually labeled result combined with conventional algorithm.
Total cone counts manually labeled by at least one engineer were 3946±172 for 0.5mm and 2441±268 for 1.0mm. Agreement between two manual labeling was 96.0±1.8% for 0.5mm and 97.4±0.2% for 1.0mm. Among the manually labeled cones by at least one engineer, 97.5±0.9% for 0.5mm and 99.4±0.1% for 1.0mm of cones were detected by our method, whereas 90.9±3.0% for 0.5mm and 95.0±0.2% for 1.0mm of cones were detected by the conventional method. This means 261±87 for 0.5mm and 107±9 for 1.0mm additional cones were labeled correctly with our algorithm. Ratio of hexagon was also improved for all six images, in average 51.6±2.3 % to 54.4±2.2% for 0.5mm and 51.8±2.7 % to 53.4±2.4% for 1.0mm.
Newly developed cone detection algorithm detected more cones correctly when some cones did not have local maximum intensity profile. As these effects were stronger in the region where the cone density was higher, improved capability of detecting cones of our algorithm was more evident at eccentricities of 0.5mm than 1.0mm.
This PDF is available to Subscribers Only