Page 146 - Computational Retinal Image Analysis

140 CHAPTER 8 Image quality assessment

and image analysis (either human or automated). Teleophthalmology systems rely on
capturing the image at a remote location and then transmitting the image to an expert
some geographical distance away. On most occasions, the expert opinion of the
diagnosis is made some time after the original image was captured. Therefore, image
quality is an issue that should be addressed at the point of capture for these systems,
allowing the opportunity for repeat image capture. Automated analysis techniques
are not common in telemedicine systems given the varying quality of the images
taken from portable devices and the relatively low-resolution of images captured
from portable devices that are suitable for transmission. However, the combination
of real-time IQA algorithms combined with advances in analysis techniques and
transmission technology is likely to lead to further developments in this area.
1.2.3 Epidemiology study requirements
In addition to large scale screening programs, population based datasets, such
as UK Biobank [24], containing tens of thousands of retinal fundus and optical
coherence tomography (OCT) images have recently become available. In addition
to imaging data, UK Biobank includes general health status (including self-reported
health), and information from routine secondary data sources (such as primary care
records and Hospital Episode Statistics), data from physical examination (including
anthropometry, measures of body composition and blood pressure) and biological
samples (including blood, urine and saliva samples) for each participant, providing a
valuable resource for the prediction and prevention of both ophthalmic and systemic
disease. Progress in linking retinal images with a range of disease biomarkers using
this large dataset have been reported [25–28]. The image quality requirements
of these systems are very different to those required for diagnostic purposes. For
epidemiological studies, the criteria are that the image clarity must be sufficient
to allow for the accurate recognition of features on the image (for example vessel
segmentation) for a sufficient portion of the image. This holds because useful
information can still be extracted from images where a portion of the image is clear,
and this can contribute towards analyses relating retinal morphometry to disease
risk and outcomes. This approach reduces wastage of images and allows maximum
information to be extracted for epidemiological analysis, since for many large
population based datasets image quality is often compromised due to avoidance of
pharmacological mydriasis to maximize participation [6].

2 Automated image quality assessment algorithms
2.1 An overview of techniques
The majority of automated image quality assessment algorithms have been developed
to determine the suitability of an image for diagnostic purposes. However, automated
IQA algorithms applied to epidemiological studies have emerged more recently. The
following brief overview will summarize IQA algorithms firstly, according to those
that are based on generic image quality parameters such as illumination, contrast