Page 71 - Computational Retinal Image Analysis

62 CHAPTER 4 Retinal image preprocessing, enhancement, and registration

3.2 Tomographic imaging
In OCT imaging, noise reduction addresses an additional source of noise, besides
conventional noise due to electronics. “Speckle noise” is an OCT artifact [56], due to
the reflective nature of the retina. So-called “speckles” are due to the interference of
the illumination with back-scattered light.
Generic, single-image noise reduction has been based on linear filtering [57],
adaptive transforms [58], wavelets [59–63], and wave atom transformations [64].
Nonlinear filtering has also been proposed, using conventional [65] or multiscale
anisotropic diffusion [66]. Other approaches include regularization [67], PCA [68],
Bayesian inference [69], and stochastic methods [70]. Compressive sensing and
sparse representations were proposed in Refs. [71–73]. A comparative evaluation of
such approaches can be found in Ref. [74].
Nevertheless, the main focus of noise reduction approaches is on speckle noise
reduction, as speckles significantly obscure retinal structure in the acquired images.
The majority of noise reduction approaches averages multiple, uncorrelated scans
of the same section. In this way, image structure due to transient speckle noise is
attenuated over structure due to actual tissue. Some techniques include adaptations
upon the conventional OCT apparatus, leading to more complex image acquisition.
In these cases, acquisition of uncorrelated scans is based on modulation of the
incidence angle of illumination [75–77], detection angle of back-scattered light
[77], laser illumination frequency [78–80], and illumination polarization [81–83].
On the other hand, spatial compounding techniques [32–34, 84] do not require
modification of the OCT scanner, as they use the purposeful motion of the scanner
to acquire overlapping and adjacent scans. Motion is a priori known minus the
uncertainty of mechanical motion and, thus, only minor alignment is required. As the
aforementioned techniques image the same tissue multiple times, they are limited
by eye motion (i.e., saccadic). Thereby, the brevity of acquisition time is required to
reduce the probability of corresponding motion artifacts.
Accidental and purposeful motions call for scan alignment, through image
registration (see Section 4.2). Once scans are registered, postprocessing further
enhances the volumetric signal. In Refs. [32, 85], 3D wavelet filters are applied to
volumetrically registered scans. In Ref. [86], volumetric neighborhoods are matched
and averaged. In Ref. [87], a physical model of speckle formation is employed and
estimated as a convex optimization problem upon the volumetric data.

4 Retinal image registration
The problem of image registration regards a test and a reference image. The goal is
the estimation of the aligning transformation that warps the test image, so that retinal
points in the warped image occur at the same pixel locations as in the reference
image. RIR is challenging due to optical differences across modalities or devices,
optical distortions due to the eye lens and vitreous humor, anatomical changes due to
lesions or disease, as well as acquisition artifacts. Viewpoint differences (i.e., due to