Page 110 - Computational Retinal Image Analysis

3 Vessel segmentation 103

A systematic approach is conducted in Ref. [52], where vessel structure is segmented
using morphological operators, based on which the OD and macular are then local-
ized by likelihood ratio tests. Mendonca et al. [53] use four directional differential
operators to detect the vessel centerlines, which then facilitate the morphological
reconstruction of vessels. The work of Miri and Mahloojifar [54] proposes to use
curvelet transform and morphology operators for edge enhancement, then apply a
simple threshold along with connected components analysis for delivering the final
segmentation. Martinez-Perez et al. [55] conduct segmentation by a combination
of multiscale analysis of gradient and Hessian information and region-growing ap-
proach. A multiresolution 2D Hermite model is investigated in Ref. [56], where a
quad-tree is used to organize the spatial configuration, an expectation- maximization
type optimization procedure is developed to estimate the model parameters.
Moreover, Bankhead et al. [57] utilize isotropic undecimated wavelet transform for
unsupervised segmentation. In segmenting UWFI Fundus FA images, Perez-Rovira
et al. [37] employ steerable filters and adaptive thresholding. An iterative approach
is devised in Ref. [58] to include new vessel structures by local adaptive thresholds.
A similar strategy is also considered by Xu et al. [59], where a tree-structured shape
space is considered and for retinal vessel segmentation, local threshold is applied
recursively. The work of Kovacs and Hajdu [60] advocates a two-step process. First,
generalized Garbor filter-based template matching is used to extract the centerlines.
Second, iterative contour reconstruction is carried out based on the intensity charac-
teristics of the vessel contours.
Variational methods have been another popular line of techniques in retinal
vascular segmentation. A deformable contour model is adopted by Espona et al.
[61], by incorporating the snake method with domain-specific knowledge such as
the topological properties of blood vessels. In Ref. [62], a dedicated active contour
model is developed that uses two pairs of contours to capture each side of the vessel
edges. To address the challenging pathological images, Lam and Yan [63] investigate
the application of divergence operator in the gradient vector field. Moreover, to deal
with the issue of multiconcavity in the intensity profile of especially pathological
fundus images, a variational approach is taken in Ref. [64] based on perceptual
transform and regularization-based techniques. An active contour model with local
morphology fitting is discussed in Ref. [65] to segment vessels in 2D angiogram.
The combined use of regional information and active contour techniques is further
considered in Refs. [66, 67]. In Ref. [35], new filters are proposed based on lifting a 2D
image is lifted by Lie-group into a 3D orientation score, and by applying multiscale
second-order Gaussian derivatives, and follow-up eigensystem analysis of the left-
invariant Hessian matrix. After projecting back from 3D space to 2D image plane, the
segmentation result is obtained by applying a global threshold. Recently, a minimal
path approach is reported in Ref. [68], where a dynamic Riemannian metric is updated
during the course of a single-pass fast marching method. It is sometimes advantageous
to perform interactive image analysis. This is addressed by Poon et al. [69],
where a multiscale filtering approach is designed to simultaneously compute center-
lines and boundaries of the retinal vessels.