Page 75 - Computational Retinal Image Analysis

66 CHAPTER 4 Retinal image preprocessing, enhancement, and registration

4.2 Tomographic imaging
Approaches to the registration of OCT images differ to fundus image registration
approaches, due to their tomographic content. In some cases, OCT image registration
is guided by the use of an additional fundus as a reference, which does not contain
motion artifacts as it is instantaneously acquired.
Noise reduction in OCT images is based on averaging registered scans. In Ref.
[32], the known translational motion of the scanner is assumed as accurate enough
to juxtapose individual scans as volumetric data. In Ref. [33], an initial manual
registration is refined through cross-correlation alignment. In Ref. [85], detecting
the retinal pigment epithelium and requiring its continuity in adjacent scans provide
a cue to registration. The work in Ref. [34] utilizes correlation maps of adjacent
images. In Ref. [135], hierarchical affine-motion estimation approach is proposed.
Low-rank and sparse image decomposition alignment has been employed in Ref.
[136].
Eye motion estimation through OCT image registration has also been used to
compensate for eye motion. In Ref. [137], an SLO-based eye tracker measured eye
motion during image acquisition and images were registered according to these
motion estimates. In Ref. [138], individual scans are registered to an SLO image.
In Ref. [139], scans are registered without the use of a reference image; a particle
filtering optimization is utilized to align scans in 3D as a dynamic system, which
is optimized when adjacent scans are in consensus. In Ref. [140], registration is
also optimization-based, but further exploits the temporal ordering of the scans and
external information on their spatial arrangement. In Ref. [141], a semiautomatic
approach is proposed to register and compare OCT scans, in order to study retinal
changes in the corresponding data.
Mosaics of OCT volumes have been formed, using a fundus image as a reference.
In Ref. [142], volumes are registered using vessel ridges and cross-correlation as a
registration cue and based on the Iterative Closest Point algorithm. In Ref. [143],
a reference image is not required, but adjacent volumes are registered based on a
B-spline free-deformation method. In Ref. [144], a six-degree-of-freedom registration
scheme for OCT volumes is proposed, which includes bundle adjustment corrections.
In Refs. [145, 146], OCT volumes are registered in 3D without any prior knowledge
of sensor or eye motion, based on correspondence of feature points.

4.3 Intramodal vs. cross-modal image registration
The combination of imaging modalities results in improved understanding of the
retinal tissue and its pathologies. Cross-modal registration enables analysis and
comparison of images that may emphasize complementary retinal features. In
tomographic imaging, registration of axial scans to a frontal, fundus image enables
registration of OCT scans, despite eye motion (see Section 4.2).
Methods in Refs. [112, 115, 116, 131] register fundus images and fluoroangiographies,
by performing vessel segmentation and matching common vessel structures across the
two cross-modal images. Vessel segmentation, detection, and matching of bifurcation