Classification of Diabetic Retinopathy by Applying an Ensemble of Architectures
Diabetes is one of the most widespread diseases with complications affecting numerous systems in the body . Every year there are millions of new patients diagnosed with diabetes. The global prevalence of the disease is also increasing with each year . Persons with a historical presence of diabetes are also prone to an aftereffect known as diabetic retinopathy. Diabetic retinopathy (DR) is a disease generated as an effect of diabetes that harms the eyes of the patient. DR happens in patients who have high sugar levels in their blood. As a consequence of diabetes, tiny blood vessels in the retina become blocked. DR is induced by the destruction of light-sensitive tissue present in blood vessels at the back of the eye. Damaged blood vessels cause a blood shortage in the region of the retina and can lead to long-lasting loss of eyesight. There is no initial sign of this disease and it can often be diagnosed only when the patient starts to lose vision. The effect of diabetic retinopathy varies from mild vision problems to complete blindness.
In persons with diabetes, the presence of retinopathy in India is about 18%. Most of the guidelines recommend annual screening of diabetic patients, even if they have no symptoms of retinopathy. Early diagnosis of this disease can be possible only through regular screening and monitoring. Proper screening of DR aids in initial detection, proper cure and a reduction in the probability of developing full blindness. Manual screening is not feasible since either patients have to come to health centres or doctors have to reach out to them. Classification of diabetic retinopathy involves assigning weight to different features and the location of these features by expert doctors. This process consumes a lot of the doctor’s time, and also sometimes that of the patient, who due to ignorance, fear or carelessness arrives at the hospital at such a late stage that it became difficult to save their eyes. Thus, computer-aided diagnosis (CAD) systems need to be developed for automatic classification of diabetic retinopathy. CAD system for the unmasking of DR takes fundus images as input. These fundus images are also used to diagnose many other eye diseases. Some of the CAD systems developed have the scope to differentiate between normal and anomalous retinal images. These systems are supposed to decrease the assigned work of the eye specialist by rejecting the healthy images with the help of the CAD system.
In this paper, an ensemble is presented that combines the results of three different standard architectures to segregate the fundus images into healthy images and images having DR. In Section 2, the work done in the space of classification of DR is discussed. In Section 3, the presented method and the dataset used are outlined. In Section 4, the outcome obtained from different architectures and ensembles are presented. Lastly, the findings are summarized in Section 5.
In the literature, different methods have been presented to classify and detect retinopathy in fundus images.
Ege et al.  presented a technique for diagnosis of diabetic retinopathy in which different statistical classifiers are tested. The tested classifiers includes Bayesian, Mahalanobis and KNN classifier. Out of the three classifiers, Mahalanobis classifier gave the best result.
Sinthanayothin et al.  discussed a method in which the principal parts of the retina are automatically detected. Then different features are extracted to identify the hard exudates, hemorrhages and microaneurysms. All the extracted features are used to conclude whether diabetic retinopathy is present or not.
Vallabha et al.  used vascular abnormalities to automatically classify diabetic retinopathy. To detect vascular abnormalities, selective Gabor filters were used. The output obtained from these features is used to classify the fundus image as moderate or severe case of diabetic retinopathy.
Sopharak et al.  presented a method to identify exudates, which is the dominant indication of diabetic retinopathy, from the digital images. Different features like standard deviation, intensity, edge pixels and hue are extracted to form a feature vector. These features are provided as input to FCM clustering method which provide detection results. These result are compared with the results provided by expert ophthalmologists.
Gulshan et al.  applied a multi-layered CNN algorithm to automatically classify diabetic retinopathy and macular edema. To detect diabetic retinopathy.
two different datasets, namely Messidor dataset and EyePACS dataset have been used. The results indicate that the presented algorithm has high specificity and sensitivity for the two diseases. The results need further validation before applying the algorithm in clinical settings.
Amin et al.  presented an automatic technique for classification and detection of exudates using fundus images. In this four main descriptors are used to choose the feature group for each candidate lesion. Combinations of different classifiers are tested and the best combination is used to enhance the performance of the system. The presented technique is compared with the existing methods using accuracy measure and area under the curve (AUC) measure. It is concluded that the suggested method outperform all the current methodologies.
Li et al.  applied transfer learning using CNN to classify retinal fundus images for diabetic retinopathy. To perform experiments two publicly available datasets namely DR1 and MESSIDOR has been used. In this different experiments have been performed in which pre-trained models are directly used as well as fine-tuning is also performed. The results proved that transfer learning improves the result as well as the performance of CNN models to classify diabetic retinopathy.
Xu et al.  explored the application of deep learning to reveal diabetic retinopathy using colored retinal fundus images. In this paper, a modified CNN network is used for the segregation and the accuracy of 94.5% is achieved.
Gargeya and Leng  proposed a deep learning algorithm that is data driven and automatically detects diabetic retinopathy. In this, 75137 publicly available fundus images are used for the purpose of testing and training. The developed model achieved 0.97 AUC using 5-fold cross verification method. The achieved performance suggests that the proposed method is highly reliable.
Akram et al.  developed a system comprising of unique hybrid classifier for the finding of lesions in the retina, which ultimately helps in the grading of diabetic retinopathy. In this system, after pre-processing, all the regions are detected which may have candidate lesions. For every contender region, a descriptor set is created using shape, intensity and statistics features. Hybrid classifier, combining Gaussian Mixture Model and m-Mediods based modeling approach, is used to enhance the grading accuracy. The architecture performance is measured using different specifiers such as accuracy, AUC, specificity and sensitivity. It achieved the AUC performance 0.981 which is superior than the individual methods.
Pratt et al.  also presented a CNN method to distinguish retinal fundus images for DR. In this, a publicly available Kaggle dataset is used and on this dataset, data augmentation is performed to enhance the medical dataset. The method achieved the validation accuracy of 75%. The images have been classified into five different classes.
Dutta et al.  proposed an automatic knowledge model in which key features are identified that can suggest the beginning of diabetic retinopathy. The proposed model is trained using three Neural networks (NN); namely back propagation NN. deep NN and convolutional NN. It is experimentally obtained that deep learning- based NN performs better than back propagation NN.
Wan et al.  presented an automatic technique to detect DR by applying CNN. The method works in three different fields of segmentation, detection and classification. Different CNN models namely AlexNet, VGGNet, GoogleNet and ResNet have been fine-tuned and analyzed. The greatest classification accuracy achieved by these models is 95.68%.
Zhao and Hamarneh  presented a method to enhance the deep learning based classifiers in analyzing retinal fundus images. The authors used prior knowledge about the vessels structures related to diabetic retinopathy. A dual layer LSTM module is also used to grasp the dependencies between different vessel structures. The authors claim that the proposed method enhances the performance up to 8%.
Rehman et al.  used a deep learning approach for diabetic retinopathy classification. The authors tested pre-trained CNN models, as well as a customized CNN model for this purpose. The best accuracy of 98.15% has been achieved on customized CNN.