Page 69 - Computational Retinal Image Analysis

60 CHAPTER 4 Retinal image preprocessing, enhancement, and registration

intensities after image formation, as well as the enhancement of the image through
noise reduction and contrast enhancement.
Image analysis methods capable of matching and aligning retinal images of the
same eye have enabled the task of retinal image registration (RIR), which also plays a
significant role in the diagnosis and the monitoring of diseases. By registering images
acquired during the same examination, higher resolution and definition images of
the retina can be obtained and enable high-precision measurements. Moreover, the
registering images into mosaics facilitates mapping wider retinal regions than a single
image does. Registration of images from different examinations facilitates follow-up
examinations and assessment of treatment through the speed of symptom reversal.

2 Intensity normalization
Intensity normalization is the form of preprocessing closest to image acquisition, as
it deals with the interpretation of pixel values for image formation. Digital sensors
have built-in intensity normalization algorithms that subsequent image processing
methods account for. Intensity normalization is also employed in the compensation
of artifacts, due to uneven illumination of retinal tissue by the imaging modality.

2.1 Fundus imaging
The interplay between optics, light source, and eye shape casts illumination of the
retina to be spatially uneven. This complicates feature detection and segmentation,
often requiring local adaptation of processing.
Contrast normalization is usually applied to the “green” channel [4–7], as it is less
sensitive to noise. Contrast enhancement in all three channels was proposed through
3D histogram equalization [8], or independent normalization of each channel [9]. In
Ref. [10], intensity is adjusted based on information from the histograms of both red
and green channels, in an attempt to reap benefits from information in both channels.
Generic, global approaches to intensity normalization have not fully solved this
problem. Zero-mean normalization and unit variance normalization compensate only
partially for lighting variation, as they introduce artifacts due to noise amplification
[11–13]. A polynomial remapping of intensities [7] exhibits similar issues.
Generic, local approaches, based on the statistics of image neighborhoods, have
been more effective and more widely adopted. To this end, Zhao et al. [14] employ
a color-constancy method. Locally adaptive histogram equalization (CLAHE) [15]
is one of the most widely used contrast normalization steps [16–23] including also
adaptations of the initial method [24–26].
Retinal imaging-specific approaches estimate a model of illumination and rectify
intensities according to the luminance it predicts [27–30]. This model is a mask
image, where pixel values are estimates of tissue reflectivity. As the illumination
source is usually unknown, this estimate is obtained assuming that local illumination
variation is smaller than across the entire image.