Diabetic Retinopathy Detection Using Deep Learning with Optimized Feature Selection

Many researchers have developed numerous automated methods for early identification of disease at initial stages but still have room for improvement in detection methods. Imani et al. in their study proposed a novel screening system for the detection of retinopathy. The author divided their work into three phases. The first phase deals with the pre-screening of the images to assess the quality of the images. The second step exploits the structural information of the images using the proposed algorithm and if the quality of the images is satisfactory then they are passed to the third phase for the detection of retinopathy. The authors implemented the MCA algorithm for the separation of retinal vessels and lesions and finally support vector machine classifier was implemented to find the normal and abnormal retinal images. For experimental purposes, the authors have used a public dataset Messidor which contains 1,200 retinal images. Their suggested system had a 92.01% sensitivity and a 95.45% specificity [16].

Gadekallu et al. [17] in their work discussed how diabetic retinopathy can be identified using a deep neural network. Feature selection was performed using principal component analysis, further author implemented the grey wolf optimization technique for optimal selection of the feature subset. They used a sigmoid activation function with an Adam optimizer for the creation of a deep neural network. The proposed network was implemented on the Debrecen dataset available at the UCI machine learning repository. The dataset has 1151 records with 20 attributes in which the last attribute shows the class label. The Messidor picture dataset served as the source for the features used in this dataset. The effectiveness of the model was evaluated in comparison to various machine learning techniques used, including decision trees, k-nearest neighbors, support vector machines, and naive Bayes. The proposed approach achieved the highest accuracy of 96%.

A novel approach using a Deep Neural network has been suggested by Wang and Yang [18]. The regression activation map was implemented by the pooling layer in the convolution neural network. The proposed model focused on the specific region of the retina image based on the severity level. Their model was able to achieve high performance as compared to most state-of-art models [18].

Forecasting the presence of Microaneurysms in the photos and detection of early-stage diabetic retinopathy is for decades still a big challenge. Prolonged elevated blood glucose, leading to microvascular problems and permanent vision loss, is inflexed upon diabetic retinopathy (DR). The retinal development of microaneurysms and macular-enema is the early symptom of retinopathy and the correct early diagnosis minimized the risk of non-proliferated DC. Quick improvements in deep learning enable effective methodology for a fascinating approach to problems of medical picture processing. Qiao et al. [19] suggested detection for the existence of Microaneurysm in fundus images by use of neural network algorithms, which integrate deep learning implemented with Graphics Processing Unit systems. A semantic segmentation algorithm was implemented on the dataset available at https://ieee-dataport.org for classifying the infected or normal fundus images. Based on their popular semanthropic, semantic segmentation divides pixels to distinguish the function of microaneurysm. This provided the medical staff with automated fundus imagery as early-moderate-extreme non-proliferated DC. The microaneurysm prognosis and early detection method for non-causal DC are recommended to effectively train the convolutional neural network for somaticizing backbone images that can improve performance and accuracy.

A learning system to classify DC grades of Fundus images has been proposed by Elswah et al. [20]. There are three phases in the current structure. Next, an amplitude normalization and raise are added to the fundus picture preprocessed. Second, to remove the compact grading function vector the pre-processed image has been entered into the ResNet Convolutionary Network (CNN) model. Finally, a stage in classification detects and decides the degree of DR, for example mild, moderate, extreme, or proliferate diabetic retinopathy (PDR). The proposed frameworks are trained on a demanding Image Dataset IDRiD. The dataset consists of 516 images which were categorized as images with DR signs and normal retinal images. To eradicate the bias of instruction, the data is calibrated such that during the exercise each DR class is shown with the same number of images. In comparison to the relevant techniques, the proposed method demonstrates an increased efficiency for the same data and with a maximum general precision of 86.67%.

In the process of the detailed field extraction from several color fundus photos, Kumar and Kumar [21] proposed an improved diabetic identification scheme for diabetic retinopathy. Diagnosis and treatment of diabetic retinopathy depend on routine eye examinations. Microaneurysms are small red retinal spots that are formed by the blood vessel's delicate inflated portion. The first step in inhibiting DR is the identification at the primary level. Various techniques for the identification and diagnosis of DR were suggested. This research proposed two characteristics, namely the number and region of MA. Pre-processing approaches such as the extraction of the green channel, histograms, and the morphological procedure were initially used. The average filtering has been used in the identification of microaneurysms, major parent review (PCA), and minimal contrasting histogram equality. Linear vector help machine classification of DR (SVM) 96% and 92% respectively observe the accuracy and precision of the DR detection system.

Alzami et al. [22] focused on stable subjects and patients with diabetic retinopathy but on extreme levels of patients with diabetic retinopathy and fractal components. By using MESSIDOR datasets and the classification of Random Forest, the author achieved the results where fractal measurements can discriminate between healthy patients and patients with diabetic retinopathy. They were not able to achieve satisfactory results in classifying the severity of diabetic retinopathy. Thus, they further explored statistical characteristics in a future direction. It must also be vigilant to spot red lesions to obtain more knowledge on the extent of diabetic retinopathy.

Deperlıoglu and Köse [23] presented a hands-on system, which included HSV, algorithm V, and histogram equalization for enhancement of retinal fundus images. The retinal fundus picture was eventually fitted with a Gaussian low-pass filter. The Convolutional Neural Network performed the classification after the image processing. In the Kaggle Diabetic Retinopathy database, a performance evaluation of the proposed procedure was carried out using 400 retinal fundus images. Every stage of image processing has been classified. After image analysis, classification research was conducted. For each point, 20 experiments were performed, and average values were established. In the experiment the precision was 97%, the sensitivity 96.67%, the specificity 93.33%, the consistency 97.78%, the reminder 93.33%, with the F-score 93.33%. The result shows that the procedure suggested is extremely effective and good at diagnosing retinal fundus images with diabetic retinopathy.

Gautam et al. [24] proposed a software-based algorithm methodology for early detection of DC, which promised to pre-detect DR without the intervention of any expert, saving expense and time. MATLAB-based image analysis is used to classify white wool lesions, cotton stains, and rough exudates associated with DR using analytical science and biomedical engineering expertise. Depending on the pixel computation counts, the eye image under investigation was categorized as a non-diabetic or diabetes retinopathy eye.

Manjramkar [25] surveyed various methods to diagnose macular enema, thus raising the value of automated macular enema evaluation. Diabetic retinopathy consists of a common community that has been observed for several years in the retina of one who has diabetes. Early visibility can be prevented by the identification of exudates.

Herliana et al. [26] presented a model for detecting diabetic retinopathy using Wide-X architecture implemented on the EyePACS dataset consisting of 5220 images. The chosen function is then further categorized using a neural network classification process. The findings of the analysis revealed the particle swarm optimization increased to 76.11%. This also validated the classification outcome improvement using 4.35 percent of the previous findings using a neural network system by just 71.76 percent.

Deperlioğlu and Köse [27] demonstrated the use of image recognition and preparation for DC using retinal fundus images. A hands-on system including HSV, algorithm V, and histogram equalization was used for the enhancement of retinal fundus images. The retinal fundus picture was eventually fitted with a Gaussian low-pass filter. The performance of the said model was assessed using the Diabetic Retinopathy dataset available publicly on Kaggle. The dataset contains 400 retinal fundus images.

Wu and Hu [28] presented a novel system for understanding transferal retinopathy. First, import dataset DIARETDB1 from the official Kaggle website, and then boost data, including data amplification, tossing, folding, and adjustment of comparison. Using pre-trained patterns like VGG19, InceptionV3, and Resnet50 etc. Finally, by the extreme stage of diabetic retinopathy, the pictures are classified into five groups. The experimental results show that the proposed approach can achieve a classification precision of 0.60, which is greater and more stable than the standard direct training method.

Chetoui et al. [29] suggest a modern, convolutional neural system named EfficientNet focused on a deeper learning architecture to diagnose re-readable diabetic retinopathy (RDRs) and vision affecting DRs. Tests have been done in EyePACS and APTOS 2019, two public datasets. The findings obtained indicate that the planned network contributes to higher speeds, hitting 0,984 RDR Area under Curve (AUC) and 0,990 EyePACS data set for vision-threatening DR. The APTOS 2019 data sets Receiver Operating Characteristics are equal to 0.966 and 0.998 respectively for a refractive and vision-proof DR. An algorithm of explain ability has already been developed and illustrates the utility of the proposal to identify DR signals.

Ekatpure and Jain [30] proposed automatic DC diagnosis by utilizing an eye fundus graphic. For the identification of red lesions, a Gaussian filter is used to suppress blood Vessels. Supporting vector machine (SVM) for red accident identification is a classifier. Red lesion characteristics for SVM training are discovered and used.

Dhivya et al. [31] detected the DR and displayed the retinal blood vessels but did not discern early stages and could not treat distinguishing features using numerous computer-based technology. The computer-based vision findings have limited precision. This research is used in the characterization of different phases of diabetic retinopathy by artificial neural networks. The findings indicate greater precision and efficiency.

Diabetic retinopathy is an impaired retinal microvasculature complication of diabetes. As retinopathy has diabetic side effects that do not induce any signs or just minor visions, Esmail and Alzain [32] have recommended the concept of an optical device with cell phone cameras and a Web-based remote medicine system for scanning diabetic retinopathy, and is very innovative in the way retinopathy has been taken to obtain retinal imaging. The web-based interface for the database management system was designed in PHP programming language and My SQL. The outcome of the retinal picture is obtained with the proposed system, and the ophthalmic specialist sends this website for further study. The suggested method provides patients with monitoring and reduces their costs by using a tracking technology.

The automatic framework for detecting and categorizing DC through the implementation of the ImageNet Model, by Jayakumari et al. [33], has been proposed. The 'Kaggle' dataset of the retinal images is available publicly for comparison and review of the algorithm's results. The ImageNet model achieved the highest accuracy in the identification of retinopathy. The model was able to achieve 98.6% of accuracy. The model has achieved noDR precision of 86.6%, modest of 62.5%, mild of 66.6%, extreme of 57.1% with PositiveDR of 42.8%. The accuracy increased by training the network in each classification with more data.

Most people with diabetic retinal disorder known as diabetic retinopathy have a medical problem. High blood sugar levels in the retina called Diabetes Mellitus in the long term are the primary cause of diabetic retinopathy. The focus of Chakrabarty [34] was on automated classification in the High-resolution Fundus picture of the Retina of patients with diabetic retinopathy and not having the same. For this reason, the images were initially processed, and mostly colored images (RGB) were translated into ideal grayscales and transformed into sizing. The processed images are then incorporated in a Deep Learning Approach, which makes a predictor about whether the patient is diabetic or not. A new technique is developed. This approach is used on a dataset of 30 high-resolution fundus retina images. The findings obtained are 100% statistical precision and 100% sensitivity. Such an Automatic Device is convenient for diabetic and stable patients to identify retina images, thereby minimizing the number of doctor feedback.

Diabetic retinopathy is a condition caused by the missing mechanism of insulin. Blood retinal arteries weakened by haemorrhage and diabetic retinopathy, exudate, and neo-vascularization are present, so vision loss or disease is observed. Two stages are studied for the improvement of the disease: proliferation and non-proliferation. For treatment, early-stage identification (NPRD) is important. Difficult exudates are the signs of early diagnosis of diabetic retinopathy. The latest approach for detecting diabetic retinopathy has been suggested by Yasar et al. [35], focused on the identification of exudates using the Kirsch protocol in retinal funds, edge detection, threshold, filters, elimination of noise, histogram equalization, optical disc removal, and morphological classification processes. This is used for precision in multiple pictures. Therefore, early diabetic retinopathy is observed in a particular manner that is based on hard exudates and hard exudates.

Nguyen et al. [36] suggested another novel approach using a deep neural network. They have implemented a regression activation map after the pooling layer in the convolution neural network. The proposed model can focus on the specific region of the retina image based on the severity level. Their model was able to achieve high performance as compared to the state-of-the-art models. Gargeya and Leng [37] in the study presented an algorithm based on a data-driven deep-learning method. For experimental purposes, the author considered 75137 fundus images which are publicly available, and got those images verified by retinal specialists for the quality of images. They further tested their model on Messidor-2 and e-optha databases. Their model achieved 94% sensitivity and 98% specificity when implemented on a local database. Another novel approach to bimodal learning was proposed by Hua et al. [38]. They implemented the bimodal learning approach for the early detection of DR that involves both baseline and follow-up information of retinal fundus images. They further implemented a deep network based on skip connection for combining different modality features. Their model achieved an accuracy of 90.6%. Gowda and Viswanatha [39] in their study suggested the use of the fuzzy c-means clustering method. They implemented the method on a dataset with 100 images for the identification of the exact region of DR. Their method was able to achieve a 99.01% accuracy rate, a 98.38% sensitivity rate, and a 96.36% specificity rate.

Vijayan et al. [40] in their study compared different ML algorithms for the early identification of the disease. For the experimental purpose, they used a retinal image dataset available at Kaggle. Feature extraction was carried out using a Gabor filter and models were constructed using random forest, k-nearest neighbor, J48, and one rule. Random forest achieved the best accuracy of 70.15% among all the methods. Another novel method with two types of ensemble classifiers was adapted by Salamat et al. [41]. They make use of fuzzy random forest and dominance-based rough set methods to detect the possibilities of a patient for diabetic retinopathy. The proposed approach achieved over 80% sensitivity and specificity.

A hybrid approach of convolutional neural network with support vector machine was suggested by Katada et al. [42]. They implemented their model on open-source American clinical fundus images and obtained a sensitivity of 81.5% and specificity of 71.9% and further evaluated the performance of their model with another dataset obtained from Keio University. The proposed model achieved a sensitivity of 90.8% and specificity of 80.0%.

This section presents the result of the experiment for identification of diabetes. Table 4 shows the result of different machine learning techniques implemented on the extracted dataset using CFS+PSO. The performance parameters are accuracy, mean absolute error (MAE), root mean squared error (RMSE), relative absolute error, AUC - ROC and F-measure.

Results obtained with CFS and PSO

Random forest achieves an accuracy of 90.4% which is the highest as compared to FRUIA and decision tree. The decision tree has the lowest accuracy i.e., 89.13. Table 5 shows the value of Precision and recall using a reduced dataset with CFS + PSO.

Table 9. Precision and recall using a reduced dataset using an optimized dataset

The hybridization also improved the performance of FURIA as well as the decision tree as compared to reduced feature subset A. An improvement of 0.9% accuracy is reported in the case of FURIA and 2% in the case of decision tree. An improvement of 1.06% accuracy is reported in the case of random forest with hybridization as compared with feature subset B. A 0.57% reduction in accuracy is reported in the case of FURIA as compared to Feature subset B and a 0.27% increase in accuracy is reported in the case of decision tree as compared to Feature subset B.

3.png

Figure 3. MAE/RMSE/ROC/F-measure on optimized data

Deep Learning Model on Optimized Dataset

We implemented a deep learning-based model on an Optimized dataset.

Table 10. Performance of deep learning-based model using an optimized dataset

The accuracy achieved by the Deep learning-based model on an optimized dataset is 93.5%, the error rate is 6.5% and the value of F1-Score is 0.94%. Figure 3 shows the comparison of MAE, RMSE, ROC and F-Measure on optimized data with deep learning model.

Figure 4 demonstrates that the Deep Learning method was able to attain the highest accuracy (93.5%) and the lowest error rate (6.5%) utilizing an optimized data set. The FURIA recorded the highest error rate of 9.27% and the lowest accuracy of 90.73%. Table 10 shows the performance of the deep learning method on an optimized dataset.

4.png

Figure 4. Accuracy and error rate of classifiers using the optimized dataset

In this article, a unique deep-learning method with an optimized feature selection for detecting diabetic retinopathy is proposed. Diabetic retinopathy, a prevalent complication of diabetes, stands as a significant cause of preventable blindness globally. The uniqueness of this research lies in its fusion of cutting-edge deep learning techniques with an astutely devised feature selection methodology. By assimilating the power of neural networks (NNs) and other deep learning architectures, the study aims to harness the inherent ability of these models to automatically learn relevant hierarchies of features from medical data. By selecting the most pertinent and instructive features from the initial dataset, the main goal was to increase the precision and effectiveness of diabetic retinopathy diagnosis. The proposed study was able to condense the feature space while retaining the most important features by combining deep learning with feature selection based on information gain. The study showed encouraging outcomes in the identification of diabetic retinopathy by utilizing deep learning models and including optimized feature selection strategies. We acknowledge that there is still an opportunity for development, though. Future studies could analyze the efficacy of various feature selection techniques and look at the effects of various deep learning architectures.