7
Detection of Diabetic Retinopathy Using Ensemble Learning Techniques
Anirban Dutta, Parul Agarwal*, Anushka Mittal, Shishir Khandelwal and Shikha Mehta
Jaypee Institute of Information Technology, Noida, India
Abstract
Reliable detection of Diabetic Retinopathy (DR) in digital fundus images still remains a challenging task in medical imaging. Early detection of DR is essential to save a patient’s loss as it is an irreversible loss. Premature symptoms of DR include the growth of features like hemorrhages, exudates, and microaneurysms on the retinal surface of the eye. The primary focus of this paper is on the detection of the severity of DR by considering the individual and the combined features. Image processing techniques are applied for the automated extraction of the retinal features. Machine learning algorithms along with ensemble learning techniques are used on these features for further detection of DR. In medical imaging, where the available data is highly imbalanced, specifically designed ensemble learning techniques are proved to be very helpful. In this paper, three ensemble learning algorithms—AdaNaive, AdaSVM, and Adaforest—are developed and compared with machine learning algorithms for binary and multi-classification of DR. Experimental results reveal that proposed algorithms outperform existing algorithms.
Keywords: Diabetic Retinopathy, machine learning, ensemble learning techniques, binary classification, multiclass classification
7.1 Introduction
As reported by international diabetes federation (IDF), in 2019 approximately 460 million adults in the age of 20–80 years are suffering from diabetes and the number is expected to rise to 800 million by 2050. Due to high blood glucose levels, diabetic people are relatively more prone to serious health problems that affect heart, eyes, kidneys, nerves, and teeth. It is also one of the major causes of blindness, cardiovascular disease, kidney failure, and lower limb amputation. Diabetic Retinopathy (DR), a retinal vascular disease is the most common cause of losing vision and blindness among working-age adults. It is caused due to the damage of the blood vessels of an individual’s light-sensitive tissue located at the back of the eye (retina) [1]. The longer a person suffers from diabetes and the less controlled the blood sugar level is, the more likely the person is to develop this eye disease. DR progresses through four stages. The first stage DR is Mild Non-proliferative Retinopathy which is characterized by the balloon-like swelling in the blood vessels of the retina. These are called microaneurysms and these vessels can leak into the eyes. The second stage is the Moderate Non-proliferative Retinopathy in which blood vessels nourishing the retina swell and may even get blocked. The third stage is the Severe Proliferative Retinopathy. In this stage, growing blood vessels may get blocked. As a result, the retina is informed to produce new blood vessels. The fourth and last stage is Proliferative DR. New blood vessels will start growing in retina, which are abnormal and delicate. These blood vessels can easily leak which causes blurred vision. Scary tissue formation causes retinal detachment and may lead to spotty vision and severe vision loss [2]. Successful classification of DR is very important to save the patient from vision loss. Therefore, accurate estimation of this disease is critical for timely prevention and care interventions.
With the exponential increase in healthcare data, interdisciplinary research is gaining huge attention among the researchers. ML algorithms are playing a leading role in discovering useful patterns from the existing health records by providing useful insights to medical practitioners which aid in the diagnosis of various disorders. Understanding and analyzing healthcare data is a sensitive issue as it deals with the life of human beings. Accuracy of prediction is very critical as false positives and false negatives both are life threatening. In literature, DR is considered as both a binary classification problem and multiclass classification problem. Binary classification only indicates whether a person is suffering from the diseases or not and multiclass classification helps in identifying the stage of DR in the patients. Thus, to predict the stage of DR, all the features need to be extracted from their specific location. Extraction of features and then further classification is a highly time-consuming task and hence requires automation [3]. A significant amount of work has been done by researchers on this automation using various ML algorithms like KNN and SVM [4–6]. In this paper, all the individual features are extracted from the retinal images using image processing techniques like Canny Edge Detection and Histogram equalization. ML and ensemble learning techniques are applied for binary and multiclass classification. This paper proposes three ensemble models for prediction of DR. The performance of these models is compared to its various existing algorithms. It is observed from experimental results that proposed ensemble techniques give better accuracy as compared to conventional algorithms. ROC curves are plotted for the same to ensure the efficacy of developed techniques. The workflow of this paper is described as follows. Section 7.2 describes the overview of the related work. Sections 7.3 and 7.4 describe the methodology and proposed system of our work. Section 7.5 describes the experimental results and analysis. Section 7.6 concludes the paper with future work.
7.2 Related Work
Table 7.1 depicts the literature of DR with the algorithm developed and limitations.
The various studies discussed above very limited feature analysis are done on extracted and combined features together for both binary and multiclass classification. Also, the accuracy achieved by algorithms on multiclass classification is less than 50%. This forms the motivation to develop new ensemble techniques.
7.3 Methodology
The proposed work is mainly divided into three phases—pre-processing, extraction, and classification of the disease. The pre-processing steps are necessary so as to make the images more uniform for classification. The dataset used is highly imbalanced, so before applying classification models, the count of all the levels should be made equal for more accurate results. Important features like exudates, hemorrhages, and microaneurysms are then extracted from these pre-processed images and passed on to the classifying algorithms [18].
7.3.1 Data Pre-Processing
Pre-processing of the images is important to remove the noise levels in the dataset as well as enhancing the image quality. Images are of varying brightness levels so all the images must be brought to a usable format for training the model. The images used had only a certain portion of useful data, with the rest being just black background. So these images are cropped and rescaled to 256*256 pixels. The dataset used is highly imbalanced and hence needs to be balanced before further use. Balancing is done by the augmentation of the images (rotating and cropping). This also improves the localization ability of the network.
Table 7.1 Literature survey of Diabetic Retinopathy.
| Year | Paper reference | Method/algorithm | Drawback | |
| 1. | 2020 | [7] |
• Ensemble-based techniques consisting random forest, KNN, decision trees, and logistic regression. • Diabetic Retinopathy dataset. • When the individual machine learning algorithms were compared with ensemble techniques, results showed that ensemble techniques give much better results than these individual algorithms. | Testing is done on a small dataset with very limited number of images. |
| 2. | 2017 | [8] |
• Bagging ensemble classifier was developed. • Approach involved feature extraction. In the first stage, important retinal features are extracted using machine learning techniques. In the second stage, these extracted features are used for further applying ensemble techniques like voting for final classification. • Results show that ensemble techniques outperform individual algorithms. | Image datasets with relatively small size were used for experiments.
Scope for more combinations of algorithms for developing ensemble approaches. |
| 3. | 2016 | [9] | • Focused on detection of bright and dark lesions for the early diagnosis of Diabetic Retinopathy. • The classification system is built using algorithms like Naive Bayes and SVM yielding considerable accuracy. | Focused only on certain types of lesions, limited features extracted.
Only limited machine learning algorithms are used which can be further extended. |
| 4. | 2016 | [10] | • Forward search and backward search method are the two proposed techniques used to select the best ensemble system. | The proposed model makes binary classification only. Feature extraction can be improved to get better accuracy. |
| 5. | 2015 | [11] | • The work uses an approach which involves features subset selection using the lasso along with ensemble learning techniques such as RUSboost and AdaBoost followed by 10-fold cross-validation | The work uses the NHANES dataset - which relies on patient questionnaires, making it prone to poor data quality. There is a scope to improve the dataset using additional clinical data. |
| 6. | 2014 | [12] |
• The work focuses on combining several base classifiers like KNN, AdaBoost, random forest, SVM into ensemble techniques with a precisely designed strategy. • This paper presents an accuracy of 90% on the MESSIDOR dataset. | The analysis and the results obtained are binary class classification. The study revolves around image processing algorithms rather than reliable machine learning techniques. |
| 7. | 2014 | [13] |
• The paper shows the relationship between the bias-variance trade-off and the feature extraction, ensemble algorithms, and post-classification processing. • Ensemble learning techniques like bagging and random forests are used and experimental results show the effectiveness of the approach followed. | Focus on feature extractions, feature selections are missing. The overall accuracy can be further improved by using a larger dataset with higher resolution fundus images. |
| 8. | 2014 | [14] | • Classification of retinopathy on the MESSIDOR dataset has been done through the analysis of classifiers such as support vector machines, K-nearest neighbors, AdaBoost, and Gaussian Mixture model. | Dataset employed for the purpose is highly imbalanced. The study revolves around whether the extracted spots are lesions or not, i.e., Binary classification. |
| 9. | 2013 | [15] | • Support vector machine classifiers are used to test whether the patients are severely affected or moderately affected and this information helps in a much clearer diagnosis of Diabetic Retinopathy. | The proposed work has been conducted on 5 images. The research should have been conducted on larger datasets. The accuracy of the proposed work has not been reported. |
| 10. | 2012 | [16] |
• Author reviews state of art ensemble techniques for class imbalanced datasets. • The ensemble techniques (bagging or boosting) combined with sampling techniques like undersampling or SMOTE gives better results and also eliminates the need for data pre-processing. | Classification algorithms are compared only on the basis of ROC curves. Some internal measures like Silhouette index, DB index can be used. |
| 11. | 2012 | [17] |
• The paper proposes an approach that involves gathering close microaneurysm candidates and applying a voting scheme on them to overcome the difficulty caused due to different algorithms extracting microaneurysms with different approaches. • A framework to build Micro Aneurysms based on the internal components of the detector has been proposed. • The Micro aneurysms detector proposed use pre-processing methods and candidate extractors. | In spite of an innovative approach, the system misclassified certain cases where serious cases of DR are present. Important factors such as the presence or lack of DR-specific lesions, the recognition of anatomical parts, quality of image are vital in a clinical setting have been ignored by the author. The proposed method has been tested only on the MESSIDOR dataset. |
7.3.2 Feature Extraction
From each image different features are then extracted by using various image processing techniques. The features extracted from each retinal image are Blood Vessels, Exudates, Hemorrhages, and Microaneurysms [19]. We are proposing a way to use a mix of traditional and modern image processing techniques followed by morphological changes to extract these features.
These individual features are being passed to the training models and for the first time, a physician will be able to determine as to which component of the eye is causing the problem and prescribe accordingly.
7.3.2.1 Exudates
To extract the exudates (as shown in Figure 7.1), CLAHE (Contrast Limited Adaptive Histogram Equalization) technique is used to process the green Channel of the image, followed by modification using Image Blur, Erosion, and Dilation to get appropriate shapes. Other modifications are done using built-in OpenCV tools [20].
7.3.2.2 Blood Vessels
To extract the Blood vessels (Figure 7.2), Canny Edge Detection technique is used against the main challenge to differentiate between actual blood vessels and small noise portions in the images. For the small noise which could be mislabeled as blood vessels due to the sensitive nature of the Canny algorithm - Histogram equalization technique, CLAHE is being used. Masking techniques are being employed to remove smaller blood vessels which are not important to DR [21].
Figure 7.1 Extraction of exudates.
Figure 7.2 Extraction of blood vessels.
7.3.2.3 Microaneurysms
To extract the microaneurysms (Figure 7.3), the first different sets of shapes that could be identified as blob are generated using the blob detection algorithms. The algorithms are then applied after applying an edge detection algorithm on the isolated green channel of the images followed by several rounds of erosion and dilation. Not all the shapes in the set are appropriate with respect to DR, so some have been left out [22].
7.3.2.4 Hemorrhages
For the extraction of hemorrhages (Figure 7.4), a large number of parameters of the blob detection algorithms are experimented on until finally arriving at a set of parameters that is able to detect the features accurately in maximum number of images. Masking is being used extensively to draw the extracted features on a black image. Canny is being used for edge detection before masking out the blood vessels [23].
Figure 7.3 Extraction of microaneurysms.
Figure 7.4 Extraction of hemorrhages.
7.3.3 Learning
After the extractions are completed, the learning models are trained on the images. Ensemble learning techniques such as AdaBoost, AdaNaive, and AdaSVM are used after splitting the dataset into training and testing sets. These ensemble methods can potentially provide better accuracy for the dataset as they employ a strong predictor which is built by repeatedly calling the weak learner on different distributions over the training samples. Furthermore, support vector machines (SVMs), K-nearest neighbors (KNN), and random forest classifiers are also used to train the model.
7.3.3.1 Support Vector Machines
SVM is a supervised learning model that analyzes the data and recognizes patterns used for further classification [16]. When provided with different training instances with each belonging to its output category, this model would classify the images into its corresponding output stage. Initially, binary classification is made where we classify the images into two stages, DR and non-DR [24].
7.3.3.2 K-Nearest Neighbors
The KNN algorithm is based upon classifying the input sample by looking at its nearest neighbors. The Euclidean distance is used to identify the nearest neighbors and the majority output class of these neighbors is assigned to the sample [25].
7.3.3.3 Random Forest
Random forests are an ensemble learning technique that takes in several decision trees at training time and outputs the result class which occurs the greatest number of times out of all the output classes from individual trees [26]. The main idea is to make a strong learner out of the weak learners and reduce overfitting by the large numbers through the introduction of a degree of randomness. It also mitigates the problem of high variance and high bias by the use of averaging. It, therefore, produces a reasonable model fast and efficiently.
7.3.3.4 AdaBoost
Ensemble methods have the ability to improve the accuracy as we combine various classifiers. The final classification result is made by combining the results obtained from different classifiers used. Further accuracy of such ensemble technique can be improved by the boosting method. One such boosting algorithm used is AdaBoost. It is a model which helps us achieve accuracy just above the random chance on a classification problem. Specifically used with the weak learners. The most common classification technique used with AdaBoost is one level decision trees. This algorithm is also the default base learner while using AdaBoost. Every instance in the training dataset is initially uniformly weighted. But for the further trainings, the instances which fail to correctly get classified are the ones which are assigned more weight. Hence, a strong classifier or a predictor is built by repeatedly calling the weak learners with different weight distributions over the training data. This is the main reason why we are able to observe an increase in the accuracy [27]. Figure 7.5 below depicts the working of general AdaBoost model.
7.3.3.5 Voting Technique
One of the simplest techniques for combining the classification results from multiple ML algorithms is voting. We pick up two or more individual models for training the dataset. Then the voting technique can be combining the models and average the predictions obtained from different sub models to give a new result. The predictions obtained from individual classifiers can be weighted, but specifying the accurate weights for the models is a difficult task. Hence, more advanced models have been built which can learn how to give the best weights to the predictions from the sub models.
Figure 7.5 Working of AdaBoost model.
7.4 Proposed Models
In general, ensemble techniques are known to improve the accuracy and computation time of any classification problem. One great example of such a technique is the AdaBoost algorithm. Previous work that has been done on the detection of DR using ensemble techniques presents the usage of standard AdaBoost technique, which by default uses decision trees as the base algorithm. In this paper, we have proposed a new technique of using different classifiers such as naive Bayes, SVM as the base classifiers for this algorithm. It is observed that this technique is highly successful in obtaining a much higher accuracy as compared to the standard technique.
7.4.1 AdaNaive
AdaNaive is an ensemble learning method for classification that is built by keeping the Naive Bayes algorithm as the base estimator while defining the parameters for the AdaBoost algorithm. Its working is similar to that of an AdaBoost classifier where the objective is to build a predictor by combining multiple poorly performing classifiers so that they can together work as a better classifier which gives much higher accuracy, the higher accuracy is due to the estimator being set to Naive Bayes instead of the None. Figure 7.6 below depicts the working of the AdaNaive model.
Figure 7.6 Working of AdaNaive model.
7.4.2 AdaSVM
In this algorithm, the base estimator used for the AdaBoost algorithm is the SVM model. AdaSVM classifier produces better results as compared to the default AdaBoost classifier. Working of AdaSVM is presented in Figure 7.7.
7.4.3 AdaForest
In this algorithm, the base estimator used for the AdaBoost algorithm is the Random Forest model. The random forest itself being an ensemble of many decision trees gives better classification accuracy for this model as compared to the standard technique where a single decision tree is used as the base classifier. Working of AdaForest is shown in Figure 7.8.
Figure 7.7 Working of AdaSVM model.
Figure 7.8 Working of AdaForest model.
7.5 Experimental Results and Analysis
7.5.1 Dataset
The images for DR Detection were obtained from the Kaggle dataset of 35,000 images with 5 output classes. Dataset consists of colored fundus images varying in size (height and width). Also, the count of images of different levels is highly non-uniform with the non-DR images accounting for 75% of the total dataset. Hence, the dataset used is highly imbalanced and at the same time contains a lot of un-interpretable images. All the images are labeled with the subject id as well as the left or right forex (1_left.jpeg means the left eye of the first patient.) Table 7.2 describes the distribution of various levels in the dataset [28].
7.5.2 Software and Hardware
The Kaggle dataset consists of about 35,000 images while we have worked on a patch of 1,500 images for training proposed algorithms like AdaSVM, AdaForest, AdaNaive, and other conventional algorithms. The computation time of these classifiers is significantly high; hence, Google Collaboratory, a Google Cloud Service with free GPU, is used. A significant reduction in execution time is seen.
Table 7.2 Retinopathy grades in the Kaggle dataset.
| Level | Condition | Count of images |
| 0 | No DR | 25,810 |
| 1 | Mild DR | 2,443 |
| 2 | Moderate DR | 5,292 |
| 3 | Severe DR | 873 |
| 4 | End Stage | 708 |
7.5.3 Results
The aim of our DR system is to grade the severity of the disease by using proposed classification algorithms. Retinal images without lesions are classified as normal and the abnormal images (with lesions) are further classified as mild, moderate, severe, and end-stage DR [19] as shown in Figure 7.9. A comparative analysis of the algorithms is done on our extracted feature set discussed in Section 7.3. Figures 7.1 to 7.4 depict the features extracted through image processing techniques.
Performance of Proposed Models for Binary Classification: Table 7.3 portrays the accuracies obtained in binary classification using proposed ensemble techniques in contrast to ML algorithms. Cumulatively, AdaNaive has provided results with high accuracies among all classifiers. AdaBoost accuracy is also very close to AdaNaive. In a feature wise binary classification, AdaNaive gives highest accuracy as compared to other learning algorithms. In blood vessels, AdaNaive accuracy is improved by 14.56%, 9%, 10%, 6%, and 10% from random forest, KNN, AdaBoost, SVM, and voting techniques, respectively. Random forest and AdaBoost have depicted similar results while classifying microaneurysms. SVM and random forest have provided an accuracy of 83% to classify hemorrhages. Random Forest and KNN have provided an accuracy of 85% in exudates classification. AdaNaive, AdaSVM, and AdaForest outperform all ML classifiers. In the case of combined features, AdaNaive and AdaSVM give accuracies of 76% and 72%, respectively. Percentage improvement in accuracy of AdaNaive from random forest, KNN, AdaBoost, SVM, and voting techniques are 20.8%, 24.2%, 2%, 13.2%, and 20%, respectively. Overall it can be deduced from Table 7.3 that AdaNaive gives highest accuracy and takes over all the other ensemble and machine learning algorithms.
Figure 7.9 Representative retinal images of DR in their order of increasing severity.
Table 7.3 Accuracy for binary classification using machine learning techniques.
| Features | Random forest | KNN | AdaBoost | SVM | Voting technique | AdaForest | AdaNaive | AdaSVM |
| Blood vessels | 78.125 | 82.000 | 81.250 | 84.375 | 81.5 | 84.400 | 89.500 | 89.200 |
| Microaneurysms | 84.375 | 81.000 | 84.375 | 81.250 | 69.4 | 79.900 | 87.300 | 85.200 |
| Hemorrhages | 83.330 | 73.300 | 80.000 | 83.300 | 68.7 | 80.100 | 86.000 | 83.000 |
| Exudates | 85.100 | 85.210 | 86.000 | 82.750 | 72.5 | 87.300 | 92.200 | 91.800 |
| Combined features | 63.300 | 61.672 | 75.000 | 67.600 | 63.2 | 69.000 | 76.500 | 71.800 |
Performance of Proposed models for Multiclass Classification: Table 7.4 illustrates the accuracies obtained in multiclass classification using proposed ensemble techniques in contrast to ML algorithms. On individual extracted features, say blood vessels, AdaNaive and AdaSVM perform similarly and provide an accuracy of 54%. Percentage improvement in accuracy of AdaNaive is 35.2%, 84.7%, 5%, and 37.2% from random forest, KNN, AdaBoost and SVM. In microaneurysms, AdaNaive and Adaforest provide similar accuracies of 56% and 55%. AdaNaive and AdaSVM provide an accuracy of approximately 58% while detecting hemorrhages. AdaForest performs well by providing an accuracy of 59% while detecting exudates. In combined features, AdaNaive gives highest accuracy as compared to other algorithms. AdaNaive accuracy is increased by 42%, 66.8%, 6.8%, 146%, and 25.6% from random forest, KNN, AdaBoost, SVM, and voting techniques.
We observe that the algorithms AdaNaive, AdaSVM, and AdaForest provide the highest accuracies in both binary and multiclass classification. For a better study and evaluation of the classifiers, we have plotted the Receiver operating characteristic curve or ROC curve comparing the performance of the algorithms in binary classification on the original images as shown in Figure 7.10. ROC curve is a probability curve and it shows how much the model is capable of distinguishing between classes. It is plotted between True Positive Rate (TPR) and False Positive Rate (FPR) where TPR is on the Y-axis and FPR is on the X-axis.
Table 7.4 Accuracy for multiclass classification using machine learning techniques.
| Features | Random Forest | KNN | AdaBoost | SVM | Voting technique | AdaForest | AdaNaive | AdaSVM |
| Blood vessels | 40.311 | 29.500 | 52.000 | 39.600 | 49.434 | 51.555 | 54.500 | 53.350 |
| Microaneurysms | 41.250 | 43.500 | 55.350 | 41.000 | 45.444 | 55.000 | 56.540 | 52.400 |
| Hemorrhages | 44.450 | 36.500 | 54.330 | 38.222 | 52.700 | 57.333 | 58.500 | 58.600 |
| Exudates | 42.500 | 40.100 | 56.250 | 39.300 | 55.500 | 59.450 | 58.750 | 57.500 |
| Combined features | 40.12 | 34.180 | 53.250 | 21.100 | 45.333 | 45.666 | 56.950 | 54.500 |
Figure 7.10 Comparison of classifiers using ROC curve (Binary classification).
The contribution of ensemble techniques in providing great results is quite significant as they outperform all the ML algorithms. Within ensemble algorithms, the proposed techniques of AdaNaive, AdaSVM, and AdaForest give the best results for both binary and multi-classification as shown in Figure 7.11 and 7.12 respectively. In multi-classification, the highest accuracy is achieved by the AdaForest algorithm for exudates. While in binary classification, AdaNaive performs the best. Also, from the results, it is quite evident that these algorithms are more effective in classifying the extracted features as compared to the combined ones. Hence, the extraction of these individual features plays a major role in the accurate classification of the disease.
Figure 7.11 Comparison of classifiers (Binary Classification).
Figure 7.12 Comparison of classifiers (Multi Classification).
7.6 Conclusion
With a humongous population of diabetic patients and the occurrence of DR among them, demand for such automated systems for DR detection is increasing. Several achievements have been made and acceptable outcome have been attained in feature extraction and detection of DR severity. Nevertheless, these results obtained are mostly on smaller datasets and are far away from the real-world applications. We have worked on a data-set that is moderate in size and classified with a labelling scheme that is more useful for clinical practice. This system extracts retinal features such as exudates, hemorrhages, and microaneurysms which help in better prediction of the severity of the disease. Various ML algorithms combined with ensemble techniques are then employed on these extracted features. We have developed three ensemble learning techniques, i.e., AdaNaive, AdaSVM, and AdaForest. These are developed by changing the base classifier of AdaBoost from Decision Tree to Naive Bayes, SVM, and Random Forest which brought significant change in terms of accuracy. AdaNaive gives better accuracy in contrast to majority of algorithms on binary and multiclass classification. AdaForest and AdaSVM give similar kind of performances to AdaNaive in multiclass classification. In the future, the dataset can be made more precise. Many images in the dataset are of poor quality and hence gave low results when extraction techniques were applied to them. Hence, these images need to be handled in a precautious manner.
References
1. Fong, D.S., Aiello, L., Gardner, T.W., King, G.L., Blankenship, G., Cavallerano, J.D. et al., Retinopathy in Diabetes. Diabetes Care, 27, suppl 1:s84–s87, 2004.
2. Health N, Exam- N. The Evolving Diabetes Burden in the United States. Ann. Intern. Med., 140, 4–10, 2004.
3. Akram, M.U., Khalid, S., Tariq, A., Khan, S.A., Azam, F., Detection and classification of retinal lesions for grading of diabetic retinopathy. Comput. Biol. Med., 45, 161–71, 2014.
4. Williamson, T.H., Gardner, G.G., Keating, D., Kirkness, C.M., Elliott, A.T., Automatic detection of diabetic retinopathy using neural networks. Investig. Ophthalmol. Vis. Sci., 37, 940–4, 1996.
5. Zohra, B.F. and Mohamed, B., Automated diagnosis of retinal images using the Support Vector Machine(SVM), Fac Des Sci Dep Informatique, USTO, Alger, 2009.
6. R.P. and P.A., Diagnosis of Diabetic Retinopathy Using Machine Learning Techniques. ICTACT J. Soft Comput., 3, 563–75, 2013, https://doi.org/10.21917/ijsc.2013.0083.
7. Saha, R., Chowdhury, A., Banerjee, S., Diabetic Retinopathy Related Lesions Detection and Classification Using Machine Learning Technology. Artificial Intelligence and Soft Computing, pp. 734–745, 2016.
8. Reddy, G. et al., a Learning model for Diabetic Retinopathy Classification. 2020 International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE), 2020.
9. S.S.K. and A.P., A Machine Learning Ensemble Classifier for Early Prediction of Diabetic Retinopathy. J. Med. Syst., 41, 12, 1–12, 2017.
10. Bhatia, K., Arora, S., Tomar, R., Diagnosis of diabetic retinopathy using machine learning classification algorithm. Proc. 2016 2nd Int. Conf. Next Gener. Comput. Technol. NGCT 2016, no. October, pp. 347–351, 2017.
11. Ogunyemi, O. and Kermah, D., Machine Learning Approaches for Detecting Diabetic Retinopathy from Clinical and Public Health Records. AMIA ... Annu. Symp. proceedings. AMIA Symp., vol. 2015, pp. 983–990, 2015.
12. Antal, B. and Hajdu, A., An ensemble-based system for automatic screening of diabetic retinopathy. Knowl.-Based Syst., 60, January, 20–27, 2014.
13. Tradeoff, B., Merentitis, A., Debes, C., Member, S., Heremans, R., Ensemble Learning in Hyperspectral Image Classification: Toward Selecting a Favorable. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 7, 4, 1089–1102, 2014.
14. Roychowdhury, S., Koozekanani, D.D., Parhi, K.K., DREAM: Diabetic Retinopathy Analysis Using Machine Learning. IEEE J. Biomed. Health Inform., 18, 5, 1717–1728, 2014.
15. Gandhi, M. and Dhanasekaran, R., Diagnosis of diabetic retinopathy using morphological process and SVM classifier. Int. Conf. Commun. Signal Process. ICCSP 2013 – Proc., pp. 873–877, 2013.
16. Galar, M., Fernandez, A., Barrenechea, E., Bustince, H., Herrera, F., A review on ensembles for the class imbalance problem: Bagging-, boosting-, and hybrid-based approaches. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev., 42, 4, 463–484, 2012.
17. Antal, B. and Hajdu, A., An ensemble-based system for microaneurysm detection and diabetic retinopathy grading. IEEE Trans. Biomed. Eng., 59, 6, 1720–1726, 2012.
18. Walter, T., Klein, J.C., Massin, P., Erginay, A., A contribution of image processing to the diagnosis of diabetic retinopathy - Detection of exudates in color fundus images of the human retina. IEEE Trans. Med. Imaging, 21, 1236–43, 2002, https://doi.org/10.1109/TMI.2002.806290.
19. Panchal, P., Bhojani, R., Panchal, T., An Algorithm for Retinal Feature Extraction Using Hybrid Approach. Proc. Comput. Sci., 79, 61–8, https://doi.org/10.1016/j.procs.2016.03.009, 2016.
20. Biran, A. and SobheBidari KR, P., Automatic Method for Exudates and Hemorrhages Detection from Fundus Retinal Images. Int. J. Comput. Electr. Autom. Control Inf. Eng., 9, 1443–6, 2016.
21. Soomro, T.A., Paul, M., Gao, J., Zheng, L., Retinal blood vessel extraction method based on basic filtering schemes. Proc - Int Conf Image Process ICIP 2018, 2017–Septe, pp. 4422–6, https://doi.org/10.1109/ICIP.2017.8297118.
22. Akram, M.U., Khalid, S., Khan, S.A., Identification and classification of microaneurysms for early detection of diabetic retinopathy. Pattern Recognit., 46, 107–16, https://doi.org/10.1016/j.patcog.2012.07.002, 2013.
23. Bae, J.P., Kim, K.G., Kang, H.C., Jeong, C.B., Park, K.H., Hwang, J.M., A study on hemorrhage detection using hybrid method in fundus images. J. Digit. Imaging, 24, 394–404, https://doi.org/10.1007/s10278-010-9274-9, 2011.
24. Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B., Support vector machines. IEEE Intell. Syst. App., 13, 18–28, 1998.
25. Zhang, M.-L. and Zhou, Z.-H., A k-nearest neighbor based algorithm for multi-label classification. IEEE Int. Conf. Granul. Comput., vol. 2, pp. 718– 721, 2005, https://doi.org/10.1109/grc.2005.1547385.
26. Liaw, A. and Wiener, M., Classification and Regression by random Forest. R News, 2, 3, 18–22, 2003.
27. Rätsch, G., Onoda, T., Müller, K.R., Soft margins for AdaBoost. Mach. Learn., 42, 287–320, 2001, https://doi.org/10.1023/A:1007618119488.
28. Diabetic Retinopathy Detection, n.d. https://www.kaggle.com/c/diabetic-retinopathy-detection/data, 2015.
- *Corresponding author: [email protected]