Page 70 - Computational Retinal Image Analysis

3 Noise reduction and contrast enhancement 61

2.2 Tomographic imaging
In OCT imaging, the need for intensity rectification stems from the automatic
intensity rescaling and averaging of the sensor, which may differ in each scan.
Generic intensity normalization has been applied to individual scans [31]. However,
prominent noise reduction methods utilize multiple scans that image the same area
and average measurements corresponding to the same tissue points [32–34]. The
central step in such approaches is the registration of scans (see Section 4.2).

3 Noise reduction and contrast enhancement
Noise reduction and contrast enhancement are typical preprocessing steps in multiple
domains of image analysis. Their goal is to improve image definition and fidelity, by
accenting image structure and reducing image noise. Generic approaches to these
tasks have been employed for all retinal image types. Nonetheless, approaches
targeted to the particular modality and the subsequent analysis have been more
widely adopted. Notably, in OCT imaging, noise reduction methods address an
imaging artifact pertinent only to this modality, besides conventional image noise.

3.1 Fundus imaging
Generic approaches to local contrast and structural adjustment have been proposed
[6, 7, 35–37], but are culpable in amplifying noise or noninteresting structures.
Single-scale [38] and multi-scale [13, 39–41] linear filterings have also been tools
of similar approaches but conversely they filter out useful fine image structure at
fine scales. Morphological transformations have also been proposed, such as the
contourlet [42] and the top-hat transformations [43], and were combined with
histogram equalization [44] and matched filters [45]. Nonlinear filtering approaches
have also been proposed. A diffusion-based “shock” filter was employed in Ref. [46].
In Ref. [47], inverse diffusion provided feature-preserving noise reduction. In Ref.
[48], sparse coding and a representation dictionary were utilized to represent vessels
and reconstruct the original image without noise.
Target-oriented contrast enhancement methods have also been tailored. Image
sharpness was amplified by compensating for the blur of the eye’s point spread
function, through the modeling of the former [49]. Multi-orientation filtering kernels
that mimic and highly correlate with vessel intensity profiles were used in Refs.
[50–52] to enhance vessels. By modeling vessel appearance as ridge structures, the
“vesselness,” or Frangi filter in Ref. [53] has been widely utilized to accent vessels
over the rest of the image.
Combination of multiple images has also been based on the information
redundancy that is provided when imaging the same tissue multiple times, using
averaging [54], or blind deconvolution [55]. The central step in these approaches is
image registration (see Section 4.2).