Page 152 - Computational Retinal Image Analysis

146 CHAPTER 8 Image quality assessment

Forward Backpropagation Neural Network with 14 neurons in the hidden layer. The
performance of the classifier was plotted on an ROC curve with an AUC = 0.9987.
The assessment of four generic image quality parameters in the algorithm of color,
focus, contrast and illumination combined with the speed of the algorithm could
provide useful information for a fundus camera operator with the aim of correction
of low quality images at the point of image capture.

2.3.2 Algorithms based on structural image quality parameters
The following algorithms show how techniques which rely on structural image
quality parameters have been applied to different applications (diabetic retinopathy
screening and population based studies).

Image structure clustering
Techniques that rely upon structural image parameters are particularly suited where
an image capture protocol is in place, such as is the case for diabetic screening
systems. The image structuring technique described by Niemeijer et al. [36] utilized
the consistency of structures within a retinal image and their relative ratios. If an
image is of low quality this image structure will be disturbed. Niemeijer et al. [36]
utilized Image Structure Clustering (ISC) to represent the structures of a retinal
image. ISC is a supervised method that enables image structure and their relative
ratios to be learnt from a set of images. The method determined the main structures
present in a set of normal quality images by the application of a number of filters to
generate a set of response vectors. The response vectors were then clustered and the
clusters corresponded broadly to anatomical features in the image such as the optic
disc. Image structures from unseen images were then compared to those found in the
training set.
The response vectors were generated by a filterbank which included various
filters at multiple scales. Given the vasculature in retinal images can have different
orientations and be located at different points in the retinal image, the filters
included in the filterbank were selected to be first and second order filters that were
rotation and translation invariant. The filters were applied with different scales to
cover the range of image structures found in retinal images. Initially, the image
structures from the training set were generated by applying the filterbank to each
pixel. Because the structures in retinal images are limited, a random sample of
pixels is adequate to produce a representative set of response vectors. K-means
clustering was applied, and five clusters were determined as optimum to give the
best classification performance [36].
36
The approach described in involved generating the features from a set of 1000
training images. The features consisted of a histogram of the ISC clustered pixels,
in addition to raw RGB histograms. A classifier was trained using selected features
and then the classifier was applied to the test set of 1000 images. A support vector
machine classifier was shown to achieve optimal performance when compared to
other classifiers. The AUC was 0.9968. Misclassifications sometimes occurred
where local image quality problems existed, since global histograms were used