Abstract
Purpose: :
Many algorithms for inter–modal or intra–modal registration of retinal image data rely on the detection of features such as vessel segments or vessel branching points in the image data. Despite recent improvements in computer performance, the reliable detection of such features is still an intensive task. In particular, most algorithms have a time complexity that does not suit the requirements of real–time processing for image streams with current frame rates and image resolutions. We propose a novel algorithm to extract vessel segments and branching points from fundus photographs and retinal angiograms.
Methods: :
To optimize the algorithm with respect to computational effectiveness, image data is evaluated only on a sparse grid. If considerable noise was present in the image data, the data was filtered first, typically with a simple but fast boxcar average kernel. It is shown that for current image resolutions, even on grid line data that is sub–sampled by a factor of five, the algorithm detects vessel segments that cross the grid lines reliably. In a subsequent step, the image is divided into two interlaced coarser grids. Every grid cell of these two grids is then evaluated for the pattern of the vessels crossing the cell borders, where any specific type of pattern is called a cell signature. The signature of a cell suggests certain vessel topography inside the cell, and a signature specific cell evaluation is used to check for the presence of the assumed topography. This leads, in particular, to rapid extraction of vessel branching points and angulated vessel segments, serving as main features for the image registration.
Results: :
Applied to a wide variety of fundus photographs and angiographic images of different quality, the algorithm provided a reliable set of extracted features for good and moderate image quality. The performance was compared to an existing feature extraction method based on vessel tracing by applying directional templates in an iterative manner, and more often than not, extracted an equal or higher number of relevant features. Execution time was strongly dependent on the necessary prefiltering, but for all cases considerably less than for the vessel tracing approach.
Conclusions: :
The described algorithm extracts specific vessel topography features from retinal image data with a reliability that is equal or better compared to vessel tracing based algorithms, at significantly less execution time.
Keywords: image processing • retina • imaging/image analysis: clinical