Abstract
Purpose:
The development of a semi-automated computerized method which enables the detection and quantification of the perifoveal capillary network (PCN) in fluorescein angiography (FA) images.
Methods:
Using the MatLab R2011a software (MathWorks®) we developed an algorithm that detects automatically the perifoveal capillary bed as depicted in FA images creating simultaneously an one-pixel-wide skeleton of it. The detection process starts after delineating manually the foveal avascular zone (FAZ) in a cropped 1500μm*1500μm subimage resulting from the original FA image and hence excluding this area of the process. This method is enhanced by appropriate filters that improve the image contrast. Thereafter the algorithm calculates the capillary density in an area with a 500μm radius and in a larger one with a 750μm radius measured from the centroid of the FAZ applying 3 different greyscale thresholds. The algorithm was applied on high resolution FA images in which the PCN was distinguishable. The analyzed FA images came from 19 subjects (age=43±14years) with normal (5 subjects) or pathologic (14 subjects) perifoveal capillary morphology. Using the statistical software SPSS (version 20, IBM SPSS Statistics) we assessed the intraclass correlation coefficient (ICC) and the Pearson’s correlation coefficient for the semi-automated algorithm and for the manual capillary detection.
Results:
ICC estimates were found to be greater than 0.9 and greater than 0.8 for all three greyscale thresholds for areas of 500 μm and 750μm radius, respectively. These results indicate that the semi-automated algorithm provides highly repeatable measurements for all three greyscale thresholds for both radii. ICC was also calculated for manual capillary detection in the same areas and was found greater than 0.9 for the mentioned radii. Pearson’s correlation coefficient for the three greyscale thresholds when comparing the manual to the semi-automated process in a 500μm radius area was calculated higher than 0.87 (p=0.00) and between 0.38 and 0.45 in a 750μm radius area (p=0.00).
Conclusions:
The described semi-automated algorithm can provide highly repeatable results that are comparable to those made by manual capillary tracing. It could serve as a potential tool for distinguishing between normal and irregular PCN morphology and for the diagnosis and monitoring of capillary abnormalities.
Keywords: 496 detection •
688 retina