Page 147 - Computational Retinal Image Analysis

2 Automated image quality assessment algorithms 141

and sharpness. Second, algorithms will be grouped according to those that are based
on structuring image quality parameters such as field definitions (e.g. the position
of the main anatomical features within the image) and vascular structure. Finally,
algorithms that combine both types of information, in addition to algorithms based
on deep learning techniques that have been described recently will be summarized.
The automated IQA algorithms that are based on generic image quality parameters
examine characteristics of the image that include the representation of illumination
and contrast information. Lee et al. [29] compared the histogram of an image with a
template intensity histogram that was obtained from a set of retinal images judged as
being of sufficient quality to produce a similarity measure to aid both evaluation of
image enhancement methods and clinical diagnosis. Lalonde et al. [30] extended this
technique by including a histogram of the edge magnitude distribution in the image
in addition to the local intensity histogram templates. Bartling et al. [31] utilized a
measurement of illumination quality and image sharpness to produce a pooled quality
indicator from non-overlapping image regions. Davis et al. [32] combined 17 features
to evaluate the quality of the image in terms of color, luminance and contrast. Pires
Dias et al. [3] used the fusion of generic image quality indicators (including color,
focus, contrast and illumination) for image quality assessment relating to diagnostic
applications. Structural parameters of the image have been considered by Usher et al.
[2] which incorporated the area of the vessel segmentation map as an image quality
metric. Fleming et al. [33] included both the area of segmented vasculature with the
macula and the field definition of the image in order to determine image quality. Hunter
et al. [34], with an application aimed at diabetic retinopathy screening, calculated the
contrast and quality of segmented vessels within the macular region and combined
this with the contrast of the foveal region compared against the retinal background.
Lowell et al. [35] proposed an algorithm based on blood vessel structure within an
automatically identified circular area around the macular. Niemeijer et al. [36] used
image structure clustering on a set of response filter vectors to characterize a set of
normal images against which other image structures were compared. Giancardo et al.
[37] used an elliptical local vessel density technique that extracted local measures of
vessel density in a method aimed to assist diabetic retinopathy screening. Welikala
[6] utilized a three dimensional feature vector related to a segmented vessel map to
match to epidemiological study requirements.
Paulus et al. [38] used a combination of generic and structural image quality
parameters to provide image quality information relating to diagnosis of ophthalmic
disease. The method combined texture metrics and image structure clustering to
cluster pixels into anatomical structures. Abdel-Hamid et al. [39] tested four different
retinal image quality assessments. The analysis investigated how a change in image
resolution can affect different image quality assessment algorithms. Algorithms that
make use of convolutional neural networks to assess image quality have recently
been described. Mahapatra et al. [40] utilized a combination of convolutional neural
network approaches with saliency values across different scales to evaluate images
in the context of diabetic retinopathy. Sun et al. [41] also used a convolutional
neural network approach and explored fine tuning of pre-trained networks. Gulshan