Overview / Usage

The number of people living with diabetes is exponentially increasing; this epidemic is observed internationally, with global prevalence of diabetes projected to climb to over 500 million by 2030 [1]. Common health complications faced by patients with diabetes include cardiovascular failure, amputation, nephropathy, and blindness [2]. The latter complication of blindness is caused by diabetic retinopathy (DR), which is the most common cause of legal blindness in the working-age population in the developed world estimated to affect over 93 million people [3]. Currently, screening and monitoring for DR is a time-consuming and manual process requiring a trained clinician to evaluate digital colored photographs of the back of the eye (retinal fundus images) to indicate the presence and stage of retinopathy. Automated image processing may reduce or eliminate the need for human grading, decrease the costs of screening, and increase its accessibility [4].
DR can be identified by the presence of lesions associated with the vascular abnormalities caused by the disease. This is con- ducted by a clinician and can be extremely resource intensive. The use of digital systems have been found to be a feasible option for mobile screening environment [5], and digital photographic imaging paved the way for future development of retinal fundus image processing. The detection of abnormal features in the eye such as exudates, microaneurysms, and haemorrhages using a trained neural network have been highly effective, and can reduce a third of the workload of a human screener, but it shows little efficacy in clinical practice as a replacement for human screening [4]. In 2015, Kaggle hosted a data science competition for identifying signs of diabetic retinopathy [1]. The goal of the competition was to push the development of automated detection systems that ideally would have clinical potential.
The aim of this research is to automate screening of retinal fundus images in grading their DR severity levels. The ultimate goal is to implement such an algorithm that could be deployed on a smartphone mobile device to provide full accessibility for diabetic patients to screen their eyes in the comfort of their own home. We implement a compact deep neural network using the principles of SqueezeNet [6], and apply it to the dataset from the Kaggle competition.

Methodology / Approach

We used the 2015 Diabetic Retinopathy Detection Kaggle Competition dataset, which was provided by EyePACS2 and consists
of 35,126 training retinal fundus images and 53,576 test images. Each image was rated by clinicians for the presence of diabetic retinopathy, using severity levels from 0 to 4. The images are a collected from different types of cameras, often with various types of artifacts present.

For screening retinal fundus images, one approach has been to use algorithms to segment retinal vessels. Retinal morphology can be applied to assessing eye diseases, including Glaucoma and Diabetes. CNNs have been applied to this problem, and achieved high levels of accuracy [8, 9, 10]. With regards to direct DR severity level classification, a number of efforts have been made using deep learning. In [11], a relatively shallow neural network was trained to classify 5 severity levels with high accuracies. In [12], a variant of the Inception-v3 [13] network, was trained to grade fundus images for diabetic retinopathy, and diabetic macular edema. The model was shown to perform well in both tasks. In [14], the use of a Weighted Kappa [15] loss function was explored, which can help incorporate information about class distributions into training.

When it comes to creating the best performing networks, often the strategy is to train very large networks with millions of parameters, however, these networks can be cumbersome to run and deploy. Our goal is to create a more compact network that can be more feasible in deployment, even on a smartphone mobile platform.

In order to create a compact network, we utilized many of the design principles from SqueezeNet [6]. SqueezeNet utilizes a number of strategies to reduce the number of network parameters, while trying to still maintain high levels of performances. Our network architecture incorporates their Fire modules and simple layer bypasses. The network layout is shown in Table 1. Dropout [19] was used after the last Fire module. Our network has a total of 807,013 trainable parameters, and can be saved in a file size of 9.7 MB without using any additional model compression techniques.

Technologies Used

The fundus images in the database contain different borders, illuminations levels, contrast, and are also inconsistent in their placement in the image. In order to provide more consistency in the training set, the black border around the fundus was segmented and cropped out using Otsu’s method [20]. The amount of black border cropped is unique for each image. Only the information sections of the fundus is kept. Each image is padded with zeros to have equal dimensions and is scaled down to 512 by 512. We additionally apply contract enhancement to each image [16]. Figure 2 shows the results of our preprocessing. During training, random combinations of rotation, flipping, scaling and cropping were used to augment images.

Our network was implemented in Tensorflow3 on a Linux machine with an Nvidia Titan X GPU. Training was done with mini batch sizes of 128 using the Adam optimizer [21] and a learning rate of 0.0001. For our final model, training was done over about 550 epochs, and took approximately 70 hours to complete.

Our overall accuracy on the test set was 83.6%. At test time, we run 5 crops of each test image through the network and average the model outputs to produce our overall classification. The crops we used were 10% crops positioned at each of the corners and one 10% crop of the center of the image. This was done to improve the invariability of the network, which provided a slight performance increase.

Rankings in the Kaggle competition were not done based strictly on the accuracy of a team’s submission. It was instead based on a calculation of Quadratic Weighted Kappa score [15]. This score is meant to represent the level of agreement between two sets of classifications, while accounting for the likelihood of agreement by chance. This score does not necessarily increase with the model accuracy, since the Kappa score accounts for the distributions of classifications. In our case, our best scoring model did not also produce our highest accuracy on the test set. The winning team, Min-Pooling [16], achieved a submission Kappa score of 0.84958. Our Kappa score was lower but still strong at 0.76543, and would have given us a rank of 20.

We were able to show that a compact network is capable of adequately conducting screening, resulting in a strong score of classification. Our system was significantly smaller and less complex than that of the top teams, making it more practical for deployment. The use of SqueezeNet demonstrate the viability of using a relatively small network to achieve strong classification scores.
Future research is still needed to be done to assess the efficacy of this system in a clinical setting and to assess its application when deployed to alternative platforms such as a smartphone device. This latter use-case would be advantageous in increasing patient compliance and patient education regarding DR.

Repository

https://github.com/mei-chen/eye-screening.git