Page 148 - Computational Retinal Image Analysis

142 CHAPTER 8 Image quality assessment

et al. [16] applied a convolutional neural network to automatically grade diabetic
retinopathy and diabetic macular oedema in retinal fundus images. The algorithm
was also assessed in terms of its performance in judging image quality.
All the algorithms described in this section aim to provide an indication of image
quality in relation to various applications. Images are either divided into distinct
classes (e.g. excellent, good etc.) or quantify image quality (usually on a scale from
0 to 1). The processing speeds of the algorithms described above is variable. Methods
that use generic image quality often have low computational requirements, while
methods that use structural parameters have higher requirements. Image quality
assessment related to epidemiological studies usually take place off-line once the
entire dataset has been captured. In applications where image quality assessments are
required at the point of capture, such as in diabetic retinopathy screening systems,
real-time operation could enhance the accuracy of automated analysis algorithms
such as the automated recognition of microaneurysms.

2.2 Datasets and metrics used to evaluate image quality

A number of datasets have been used to evaluate IQA algorithms, which have been
described in the literature. In this section, we describe some of the main datasets used
along with an indication of their size. Pires Dias et al. [3] used the public datasets of
DRIVE [42], Messidor [17], ROC [43] and STARE [44] and two proprietary datasets
(it is reported that access is available on request) from DR screening initiatives in the
central region of Portugal and Channai, India. The images from Portugal were graded
by a human grader from the Association for Innovation and Biomedical Research
on Light (AIBILI). An evaluation of the method was performed on a total of 2032
retinal images. Fleming et al. [33] used a private dataset of 1039 retinal images
from 489 patients attending a diabetic retinopathy screening program. The images
were graded by a clinician for image clarity and field of view. Hunter et al. [34],
used a private dataset of two hundred retinal fundus images randomly drawn from
a diabetic retinopathy screening dataset which were accompanied by a 1–5 scale
of image quality graded by an ophthalmologist, where low scores were associated
with poorer quality. Niemeijer et al. [36] used a total of 2000 images from a dataset
obtained from a diabetic retinopathy screening program in the Netherlands. Each
image was classified by an ophthalmologist to give an image quality category on a
scale from one to four. Giancardo et al. [37] used a dataset of 84 macula centered
fundus images. They were divided into four classes: “Good”, “Fair”, “Poor” and
“Outliers”. The first 3 classes were composed of a subset of a fundus images [11, 45].
Welikala et al. [6] utilized the UK Biobank dataset, which is a large scale population
based dataset, with access via an application process. The dataset included 68,151
participants. For evaluation of image quality 800 images were used from 400 random
participants. Images were defined by an experienced retinal grader as adequate or
inadequate. Paulus et al. [38] used a private dataset of 301 retinal images. Three
human observers, including one eye expert, decided on the quality of each image
according to set criteria and they judged the image quality to be adequate for a