Vascular Tree Extraction
A segmentation of the vascular tree can be useful for diagnosis purpose
as well as it seems to be the most appropriate representation for the
image registration applications due to the three following
reasons: 1) it maps the whole retina; 2) it does not move except
in very few diseases; 3) it contains enough information for the
localization of some anchor points. Recently, we have published papers on
locating and extracting blood vessels from a retinal image.
The proposed algorithm
is composed of four steps. First, since blood vessels have lower
reflectance compared to other retinal surfaces, we apply the
matched filter to enhance blood vessels. Second, an entropy-based
thresholding scheme can be used to distinguish between
enhanced vessel segments and the background in the match-filtering
image. Subsequently, a
length filtering technique is used to remove misclassified pixels.
After that, a morphological-based thinning method is employed. Finally, vascular
crossover/bifurcation detection is performed by window-based
probing process. These vascular
crossover/bifurcation points will be used as our feature points in
the registration process.
pdf version of the papers can be downloaded here:
ETDRS Seven-Field Retinal Image Registration
This paper presents a new retinal image registration approach for
Early Treatment Diabetic Retinopathy Study (ETDRS) standard
images. The ETDRS imaging standard specifies seven thirty-degree
fields of each retina and presents three major challenges for
image registration. First, only ten to twenty five percent of
adjacent fields overlap, which leads to inadequate landmark points
(crossovers and bifurcations) for feature-based registration
methods. Second, the contrast and intensity distributions of
images are not spatially uniform or consistent. This can
deteriorate the performance of area-based registration techniques.
Third, high-resolution retinal images contain large homogeneous
nonvascular regions which result in difficulties for both
feature-based and area-based techniques. In this work, we propose
a combined area-based and feature-based retinal image registration
approach for ETDRS standard images that takes advantage of
area-based and feature-based methods. Three major steps are
involved. First, as the most prominent anatomical structure, the
vascular tree is extracted from retinal images using our recently
proposed technique. Next, zeroth-order translation is estimated by
maximizing mutual information based on the binary image pair
(area-based). Specifically, a local entropy-based peak selection
scheme and a multi-resolution searching strategy are developed to
improve accuracy and efficiency of translation estimation. Third,
we use two types of features, landmark points and sampling points,
for first and second order transformation estimation (affine and
quadratic models). Sampling points, which are acquired by imposing
a grid onto the thinned vascular tree, are only involved when
landmark points do not meet certain criteria. Simulation results
on 504 pairs of ETDRS retinal images show the effectiveness and
robustness of the proposed algorithm.
pdf version of the papers can be downloaded here:
Hybrid Retinal Image Registration,
to appear in IEEE Trans. on Information Technology in Biomedicine.
