Abstract
An automated algorithm for identifying cone photoreceptors was implemented in Matlab for aiding the analysis of adaptive optics (AO) retinal images. The algorithm exploits the same optical qualities of cones that make AO retinal imaging possible. Its performance was tested on six cropped images that have already been analyzed by the authors. Out of a total of 2,153 manually labeled cones, the algorithm correctly identified 94.1 percent of them (92.7 to 96.2 percent across the six images) with false positives ranging from 1.2 to 9.1 percent. We analyzed four large AO montages from one monkey and three human retinas acquired from 0.10 to 1.86, 0.60 to 2.60, 1.10 to 4.27, and 0.35 to 1.85 degrees eccentricity respectively. Cone densities of these montages ranged from 25,300 to 72,200 cones/mm2. Voronoi analysis was used to analyze cone packing structure. Approximately 50 percent of the cones are hexagonally packed for all images. Percentages ranged from 35 to as high as 70 percent locally at certain eccentricities. The consistency of our measurements demonstrates the reliability and practicality of an automated cone identification routine. This algorithm in Matlab function format can be downloaded from http://vision.berkeley.edu/roordalab/Kaccie/KaccieResearch.htm.
Supported by NIH Bioengineering Research Partnership EY014375 and T32 EY 07043. We thank Pavan Tiruveedhula for programming assistance. We would also like to acknowledge Joe Carroll and Curt Vogel for their comments concerning the algorithm and analysis process.