Page 153 - Computational Retinal Image Analysis

2 Automated image quality assessment algorithms 147

as features in the system. The training and test set images were derived from a
DR screening program in the Netherlands. The images were graded as either low
quality or normal by ophthalmologists. An image was graded as low quality if the
ophthalmologist felt a “reliable judgement about the absence or presence of DR in
the image” was not possible. To compare the performance of the automatic method
with a second observer, an ophthalmologist divided the images into four categories
of normal quality (definitely or possibly) and low quality (definitely or possibly).
The image structuring method above used image structure parameters similar to [33]
and [35]. However, the method did not require the segmentation of the vasculature
and other anatomical features. This makes the method generalizable and may be
useful therefore to extend to other medical imaging applications.

Segmentation map feature analysis
An approach utilizing support vector machines to classify information from
segmentation maps was employed by Welikala et al. [6] to ensure retinal image
quality assessment for epidemiological study requirements. The requirements
for inclusion of retinal images into epidemiological studies are different to those
required by DR screening programs. For epidemiological studies, maximizing
the amount of data for analysis is key to maximizing statistical power to examine
morphometric associations with disease risk and outcome. In Welikala et al. [6]
retinal vessel morphometric measurements were analyzed from the UK Biobank
[24] fundus image dataset with the aim of understanding the link between retinal
vessel morphology and cardiovascular disease risk. Within the UK Biobank dataset,
image quality varies significantly across the dataset and therefore a method for
assessing the image quality automatically for each image was important to allow
for in the analyses. The IQA algorithm developed examined the segmentation map
as an indicator of image quality. Criteria for a reliable segmentation in terms of
the epidemiological study requirements was defined in terms of three different
factors: (i) more than half the vasculature should be segmented (ii) segmentation
should not be considerably fragmented/unconnected, and (iii) non-vessel objects
should not be segmented (e.g. choroidal vessels, hemorrhages, light reflexes, etc.).
Fig. 3 shows the effect of poor image quality (poor illumination in this case) on
segmentation quality.
Features were selected to reflect these criteria which summarized the vessel
maps in terms of quantitative measures of area, fragmentation and complexity.
An SVM was used to divide the retinal images into two classes of “adequate” or
“inadequate” quality.
The algorithm was developed to ensure the accurate processing of UK Biobank
images and is included in a software system to measure vessel morphometry [52].
The UK Biobank image dataset currently contains images from near 70,000 adults
who underwent retinal imaging. A subset of 800 images drawn randomly from this
large prospective study was used to train and test the algorithm. An ophthalmic
grader graded the quality of the images with respect to the criteria listed above. The
performance of the algorithm was reported as achieving a sensitivity of 95.33%