Page 157 - Computational Retinal Image Analysis

3 Conclusion 151

towards rapid assessment of image quality to inform whether further images are
needed while the patient is still in situ.

3 Conclusion

This chapter has focused on IQA of retinal fundus camera images, given that this
method of imaging is currently most widespread in clinical applications. Retinal
fundus camera imaging provides a two-dimensional image of the eye and white light
combined with a specialized microscope attached to a camera is used to acquire the
image. Different angles of view of the retina can be taken with varying magnification.
The resolution of the retina in fundus images lies in the region of 7–20 μm [27].
Other methods of imaging the retina also require assessment of image quality.
Optical Coherence Tomography (OCT) is an important and increasingly widespread
retinal imaging method. Near infra-red light combined with interferometry are used
to generate a three-dimensional view of the layers of the retina with a resolution in
the region of 4 microns [27]. Consensus criteria related to image quality assessment
of OCT images has been explored [56] in relation to the application of identification
of imaging biomarkers for neurodegeneration in multiple sclerosis. Scanning Laser
Ophthalmoscopy (SLO) is a further principal method for imaging the retina and uses
a scanning focused laser beam to generate two-dimensional views of the retina at one
or two wavelengths by collecting the light through a confocal pinhole. Various fields
of view are available (including wide-field). Typical resolutions are in the region of
10–15 microns [27]. Image quality is reported as affecting landmark detection in the
analysis of SLO images [57].
The work described in this chapter has focused on image quality assessment
algorithms applied to retinal images taken from adult participants. Variations in
image quality resulting from prominent arteriole light reflexes in retinal fundus
camera images taken from school-age children has been reported with respect to
vascular morphometric analysis [58]. Pre-term infants require regular retinal images
to be taken to screen for ROP. Swanson reported assessment of quality [59] and IQA
was reported for ROP images [60].
In this chapter, we have demonstrated how retinal fundus images can vary in quality
according to a number of issues relating to image capture protocol and the condition
of the subject. We have shown how the assessment of image quality is dependent
upon the application, and how automated IQA algorithms can be evaluated in terms of
human judgment of quality relating to how the image will be used for different clinical/
epidemiological purposes. Applications covering screening, teleophthalmology
and epidemiology studies demonstrate the range of IQA requirements necessary
for different applications. A review of different IQA algorithms is given. Following
an overview of the field, the main public and private datasets used are summarized
together with an account of the metrics used to evaluate IQA algorithms. Finally, to
demonstrate the main methodologies that have been applied to different applications,
selected examples of IQA assessment algorithms are discussed in further detail.