Abstract
Purpose: :
To create an automated system for evaluating image quality, including alignment according to ETDRS protocol.
Methods: :
One hundred images were selected from a database of 500 patients (1000 eyes). A reference image was selected based on a visual interpretation of acceptable alignment of ETDRS field 2. Each of the remaining 99 images was compared to the reference image using two metrics: cross-correlation and mutual information. These two metrics were compared to a calculated difference in optic disc location for each image to the reference image. A function that estimates disc alignment from either metric was calculated using polynomial curve fitting. This analysis was done in two different ways. First, it was applied to images that were not manipulated. Second, it was applied to images for which the histogram was matched to that of the standard image.
Results: :
The normalized cross-correlation between images showed an inverse relationship with optic disc distance (R2=0.7) in the red channel. The mutual information metric also showed an inverse relationship with optic disc distance (R2=0.65) in the green channel with histogram equalization. Higher errors were produced by images with uneven illumination.
Conclusions: :
Finding reliable metrics to match images is essential in being able to index a set of standard images by objective measures of content. Our results show that it is feasible to build a catalog of image alignment that can be mined automatically using metrics of information content. Uses include stand alone, real time, assessment of image quality and automatic building of reference datasets.
Keywords: computational modeling • imaging/image analysis: non-clinical • retina