Page 155 - Computational Retinal Image Analysis

2 Automated image quality assessment algorithms 149

2.4 Algorithms based on deep learning
Deep learning is emerging as a leading method in the automated assessment of
retinal images with respect to diabetic retinopathy detection [16] and cardiovascular
risk [28]. Deep learning methods have emerged recently in the area of retinal image
quality assessment.

2.4.1 Convolutional neural networks
Gulshan et al. [16] applied a deep learning method to automatically detect diabetic
retinopathy (DR) and diabetic macular oedema (DMO) in retinal fundus images. The
algorithm was evaluated for its performance in detecting these two diseases and in
addition, the algorithm’s performance was also evaluated for predicting gradable and
ungradable images.
Previous traditional machine learning methods discussed in earlier methods
require features to be selected as input to various classifiers such as SVM, ensemble
etc. Deep learning algorithms require a neural network classifier with many (deep)
layers to be trained, but they do not require features to be selected before training.
The neural network takes the intensities of pixels from a large set of fundus images
as its training data input and gradually adjusts parameters during the training process
until a complex mathematical function is optimized so that the neural network
prediction matches the expert grader assessment as closely as possible. Once the
training phase is complete, the trained algorithm can be applied to assess diabetic
retinopathy severity on unseen images.
The method utilized a convolutional neural network (CNN) that was trained
for image classification relating to detection of diabetic retinopathy and diabetic
macular oedema in retinal fundus images. The algorithm was trained on a dataset
of 128,175 pre-processed retinal images. The training and test set of images were
graded for the presence of DR and DMO in addition to image quality by ophthalmic
experts.
The CNN architecture used was based on work by Szegedy et al. [54] who
used inception-v3 architecture. It was reported that distributed stochastic gradient
descent was used to train the network. Batch normalization was also utilized.
The network was trained on a variety of images that included mydriatic and
nonmydriatic images. The network was trained to make multiple binary predictions
that included severity of DR and DMO. Training on 80% of the data was used to
optimize the network weights and 20% was used to optimize hyperparameters. An
ensemble of 10 networks with the same training data was combined to produce a
final prediction.
The evaluation of the algorithm with respect to detection of DR and DMO was
reported on two datasets consisting of 9963 images (EyePACS) and 1748 images
(Messidor). At an operating point selected for high sensitivity for detecting diabetic
retinopathy or macular oedema, sensitivities of 97.5% and 96.1%, and specificities of
93.4% and 93.9% were reported in the 2 validation sets, respectively.
The evaluation of the algorithm with respect to gradable and ungradable
images was reported. The algorithm was evaluated for accuracy in assessing image