Page 156 - Computational Retinal Image Analysis

150 CHAPTER 8 Image quality assessment

quality for a subset of the data of the main study. The Messidor data set was not
included as the number of ungradable images were low. Analysis was reported for
the EyePACS-1 dataset. Eight ophthalmologists were involved in grading the 9963
EyePACS-1 images using a grading tool that recorded overall image quality. The
image quality was assessed according to criteria relating to focus, illumination, image
field definition and artifacts. For fully gradable images the sensitivity and specificity
were reported as 93.9% and 90.9%, respectively. AUC was reported as 0.978.
Additionally, the algorithm performance in detecting diabetic retinopathy was also
assessed in terms of its performance in handling mydriatic and non-mydriatic images
in the EyePACS-1 dataset, which may affect image quality. The authors report similar
performance for mydriatic images compared to non-mydriatic images and these
findings could impact screening protocols with respect to routine use of mydriasis.

2.4.2 Human visual system information combined with convolutional
neural networks
A further implementation of deep learning was reported by Mahapatra et al. [40] that
was aimed specifically at providing retinal image quality assessment and taking into
account the role of the human visual system (HVS) by the inclusion of saliency map
information. Saliency maps aim to define where a particular region is different (or
noticeable) from its neighbors with respect to image features. Itti et al. [55] previously
developed a visual attention system to select conspicuous locations of a scene. The
saliency model proposed combines multiscale feature maps to form a local saliency
map which takes both local and global features into account that are relevant for
image quality assessment and the identification of gradable and ungradable images.
In addition to information from the saliency maps, information from a trained
convolutional neural network (CNN) was included in the final image quality
assessment system. The CNN included five convolution layers and five max pooling
layers. Rectified Linear Units were used to facilitate faster training and to reduce
sensitivity to the scale of the input. During training negative log-likelihood was used
as the loss function and stochastic gradient descent using dropout was implemented
to reduce overfitting. Data augmentation was applied to expand the training set by
image translation and horizontal reflections to expand the dataset size by 50 times.
The training data comprised 80% of the total dataset size.
To produce the image quality classification, feature vectors derived from the saliency
maps were used to train one random forest (RF) classifier. A second feature vector
derived from the last fully connected layer of the CNN was used to train a second RF
classifier. Both classifiers used the same image labels. A combination of the probability
outputs from the two RF classifiers was used to give a final classification of “gradable”
or “ungradable”, therefore combining both supervised and unsupervised image features.
The algorithm was evaluated against three other techniques and the data set was
acquired from a DR screening program and included 9653 ungradable images and
11,347 gradable images. All images were non-mydriatic and were graded by human
graders. The sensitivity was reported as 98.2%, with a specificity of 97.8%. The
system was reported to have low computation time and therefore could contribute