1. INTRODUCTION:

Current research in medical field has a wide range of application in analyzing the medical images to diagnose the disease at an early stage or its severity using machine learning algorithms. For image based diagnosis, prognosis and risk assessment can be done using machine learning algorithms which consumes less time with reasonable accuracy¹. Still there is a shortage in experienced clinicians and it is required to develop new automated methods which may suggest the early detection of the disease with accurate results and manual detection is error prone some times². Diabetic Retinopathy is a growing threat to vision loss and sometimes it may lead to permanent blindness in the people who are suffering with diabetes for a long time.

This disease is associated with significant vision loss often involving long term disability and costly treatment. So prescreening of retinal images plays a major role to reduce the disease severity with proper medication. As per the statistics of World Health Organization prevalence of DR is rising with age group 40 to 60 globally. Fundus images and Optical Coherence Tomography (OCT) images are mainly used by ophthalmologists to diagnose the health of retina³. Automatic segmentation of abnormalities in DR images could help to predict the disease at an early stage and enables the usage of proper clinical procedures to reduce the vision loss. Traditional methods require a large number samples and high level of expertise to diagnose the disease an early stage. Fundus camera is the major tool used to obtain the image of retina. From the images, ophthalmologists try to observe various abnormalities like irregular vessel structures, bright spots or red lesions, and more number of dark spots near the fovea⁴. Density of microneurysms and expansion of vessel structure or hemorrhages, hard exudates and leakage of lipids in the eye near fovea are the main observations to detect the DR. Over a period of time, if DR is not diagnosed properly then it leads to Proliferative Diabetic Retinopathy⁵. Making annotations of abnormalities in fundus image for a raw data with large number of samples is ineffective and time consuming process for medical experts. Recently deep neural network learning methods are gaining popularity to detect DR at early stages^6,7. In this paper, to extract features from the Region of Interest (ROI) Alex Net DNN based model is used for feature extraction in DR images. To assist clinicians, this paper proposes an enhanced frame work of semi supervised algorithm and uses unlabelled data with modified Random forest classifier to uplift the performance. Deep Neural Networks constitutes many layers and produce thousands of parameters necessitates optimization to reduce complexity.

In this proposed work, we developed a new feature vectors for intensity information, local features and edge detectors to observe structural and statistical variations with Gaussian mixture model, Gabor filter coefficients and GLCM features for various abnormalities. In medical images each pixel has its own importance, so bright preservation and contrast enhancement is achieved with recursive weighted histogram equalization is applied at preprocessing stage. Segmentation of blood vessels is done through morphological operators and edge features are obtained through Gaussian filter, which discovers the intensity variations due to poor illumination and noise. Based on the features extracted from different abnormalities, most prominent features are selected through genetic algorithm. Fig.(i) shows the details of abnormalities in fundus images with DR. The details of severity scale as per International Clinical Diabetic Retinopathy (ICDR) to diagnose Diabetic Retinopathy mentioned in Table.(i).

Fig 1. Fundus images of Normal retinal image and Retinal image with retinal abnormalities due to Diabetes. (Image source: Image from Eye Care of Columbia for Diabetic Eye disease)

Pre-processing steps include green channel extraction, downscaling and contrast enhancement making the image suitable for further processing. The accuracy of segmentation primarily relies on the consistency of the contrast of the entire image. Each pixel has its own importance for medical image analysis, so without disturbing the image details contrast has to be increased and brightness of the image has to be preserved. Many techniques have been proposed for enhancing low contrast images. Histogram Equalization has its simplicity and good performance in almost all types of images. Histogram Equalization is classified into two classes: Global Histogram Equalization (GHE) and Local Histogram Equalization (LHE) techniques are used to detect lesions of blood vessels and the presence of optic nerve in the image. This paper evaluates the proposed method with a database which consist publicly available data sets like DRIVE, STARE and DIARET DB1. In contrast to the existing work in literature, we propose an enhanced approach to multilevel classification to detect the disease severity with segmenting the Region of Interest (ROI) from the fundus images. To examine the efficiency of the proposed framework classification has been performed with Random Forest Classifier with optimal feature selection.

This paper is organized as follows. Section 1 describes literature survey, section 2 reveals methodology to detect segmentation of various abnormalities, and feature extraction and classification procedure. Section 3 gives information about experimented results with various performance metrics and concludes with a comparative analysis of other classifiers.

Table 1. Depicts the standard DR severity scale

Severity Scale	Observations	Definition
0	No abnormalities	No DR
1	Small number of microneurysms.	Mild NPDR
2	Small number of microneurysms and retinal hemorrhages with or without cotton wool spots	Moderate NPDR
3	Numerous hemorrhages, microneurysms, micro vascular abnormalities and cotton wool spots	Severe NPDR
4	Fragile blood vessels , dense microneurysms and hard exudates.	High Risk PDR

2. METHODOLOGY:

2.1 Enhanced Frame work of Automatic Diabetic Retinopathy Detection:

This section demonstrates the details of DR classification mechanism based on severity scale. This mechanism consists of mainly four principle stages, 1) Preprocessing for brightness preservation with contrast enhancement 2) Optical Disc removal 3) segmentation procedure to segment various abnormalities like blood vessels, microneurysms and hard exudates 4) Feature map construction for selecting prominent features 5) classify the severity scale based on optimal feature selection. Firstly input image is resized and preprocessed through histogram equalization, and then Optical disk removed using Circular Hough Transform and segmented for abnormalities by applying morphological operations for blood vessels and hard Exudates. Then various features like intensity shape features are extracted using Gabor filter and edge detectors to construct feature map termed as Gradient Descent Pattern (GDP) and other GLCM features. Feature map consist intensity information, edge and shape information from GDP and GLCM features of the image. Then this information is optimized to get prominent features and further applied to classifier to classify the disease severity scale. Fig. (ii) Shows enhanced frame work of Automated Diabetic Retinopathy Diagnosis.

Fig. 2. Enhanced Frame work of Automatic Diabetic Retinopathy Diagnosis.

Preprocessing:

In preprocessing stage, given fundus image resized to 256 X 256 and converted to grey level image then subjected to Histogram equalization, Contrast Limited Adaptive Histogram Enhancement (CLAHE)^8,9, and Recursive Separated Weighted Histogram Equalization (RSWHE)^10,11. In general histogram processing is used for contrast enhancement of poor quality images and it does not preserve brightness information. In medical image analysis information about each pixel is needed, so in this approach maximum value of image is observed, minimum and maximum mean values from the histograms of luma component are extracted from the image. Mean values are rounded to nearest values in a repetitive manner. Weighting constraints are applied on the input histogram and modified histogram is equalized. This preprocessed image is represented as Im_a^H_. Enhancement calculations are adjusted based on the eq (1).

Im_a^H = Eq (1)

Segmentation:

In segmentation stage first step is removal of optical disc¹². Circular Hough Transform which uses canny edge detector to detect edges of the optical disc, and local maxima is obtained to get the center of the disc in a recursive manner CHT is immune to noise and sensitive to observe the differences between feature specification descriptions¹³. The description of CHT is given by eq(2) and this image is represented as Im_b^OR, where (a,b) is the center of the circle and (x,y ) is a fixed 2D point in the image.

Im_b^OR = Eq (2)

Blood vessel segmentation approach¹⁴follows different steps namely applying RSWHE to improve contrast of low contrast images along with the preservation of brightness then subjected to average filtering, image subtraction, grey level thresholding followed by the application of morphological operators like iterative closing and area opening functions. Gaussian filtering is applied to this image to have a stabilized pattern of blood vessels¹⁵. This image is represented as Im_c^BV. By observing maximum grey threshold, the image is subjected to binarisation. Image subtraction is performed between blood vessel removed image and binary image. The following Fig.(iv) shows the resultant images after segmentation of blood vessels, exudates^16-18 and microneurysms^19-21 and these segmented images are represented as Im_c^BV, Im_d^EX. These images are subjected to feature extraction.

Feature Extraction and Classification:

Feature extraction is a segment based approach to classify images based on textural information like shape, location of the segment etc in a given image and intensity based information²². In medical image analysis feature extraction place a major role in disease identification at an early stage for proper medication. Here in this paper DR detection is mainly done by selecting the location and shape of abnormality in the fundus images with their intensity levels and GLCM features gives special features between the pixels and textural information²³. Feature extraction is done using Gaussian filter for edge feature information, Gabor filtering is to locate the abnormality and GLCM features for obtaining statistical information of the given diabetic retinopathy images Im_a^H, Im_b^OR, Im_d^EX .

Feature map for Abnormalities:

Diabetic retinopathy image with abnormalities segmented images are Im_c^BV, Im_d^EXused to obtain informative feature extraction. The descriptor is expected to capture intensity oriented information to observe the bright lesions which are also termed as exudates. Here two filters (Gaussian and Gabor) are applied to the input image and the main intention is to highlight edges of the exudates. Since edges and their location in the input image are important to detect the disease. Gaussian feature map is constructed by utilizing Center Surround (CS) theory for human visual system. The prominent feature of the CS theory ensures enhancement of edges which ensures the detection, location and tracking of small objects. Gabor filter mainly focuses on orientation features which play major role to describe the size and location of the abnormality from the optical disc. These features provide information about texture and discrimination between pixels with abnormal vessel growth and microaneurysms²⁴. Mathematical representation of final Gaussian maps are shown below in Eq(3) Where Ga_F is the final Gaussian edge feature map, FM_a(x,y) is the map obtained from subtraction of image pixel by pixel at different levels.

Ga_F(x,y) = max(FM_a(x,y), FM_b(x,y)) Eq (3)

The necessary mathematical equations for Gabor filter is represented as

GB(x,y) = Eq (4)

Eq (5)

GLCM features observe the spatial information between the pixels of the DR images and reveal the texture information regarding the abnormality^25,26. The main features for medical images analysis is explained by Hera lick which includes entropy, contrast, homogeneity, correlation, mean and standard deviation etc. Gradient Descent Pattern (GDP) is constructed with concatenation of Ga_F(x,y), GB(x,y) and GLCM features, denoted as Gradient Descent Pattern Future Map (GDP-FM). All these features represent the spatial information between the pixels of the DR images and reveal the texture and directional information regarding the abnormality.

Classification and Optimization:

Medical image classification is growing research area which plays a prominent role in disease diagnosis. Many classifiers are deployed which includes Decision Tree, SVM and Random forest available in literature for classification problems with machine learning^27-28. Deep Learning with Alex Net is gaining popularity in disease classification with medical image analysis^29-31. Still there is a need to focus for selection and optimization of relevant features from the diseased images. In this paper Random Forest Classifier is used to discriminate the images based on two categories along with feature vectors. Optimization is process in machine learning through training the model iteratively to evaluate minimum and maximum functions thereby to increase the accuracy³². Optimization in machine learning can be observed through network model optimization, parameter optimization and feature optimization. Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are well known algorithms for feature optimization. In this work features are optimized using Jaya based algorithm^33-34.

RESULTS AND DISCUSSIONS:

The proposed approach with gradient features for diabetic retinopathy diagnosis is experimented using MATLAB 15.b by developing a data base with globally available datasets like DRIVE, STARE and DIARET DB1. Our enhanced method is experimented and evaluated using Alex Net based DNN with Random forest Classifier. The comparative analysis is performed with state of art literature. Variable population sizes of image data of 25 to 300 with labels are used to train the classifier. Input to the Random Forest classifier is feature vectors of different images, and training data set consists of data points with multiple labels and each one is given to one decision tree. During the training phase, each tree produces a prediction based on the maximum votes of the data point, each time votes are updated recursively. A novel image coding scheme with Gaussian and Gabor filter Gradients with Random Forest Classifier are used for feature optimization. In this paper for brightness preservation and contrast enhancement is achieved through recursive separated and weighted histogram equalization method is implemented. RSWHE gives 2.5% more Peak Signal to Noise Ratio, when compared to other histogram equalization techniques. By extracting the ROI (region with abnormalities) in the fundus image the proposed CNN is applied for feature extraction.

In contrast to the conventional GMM model this paper uses gradient pattern with concatenation of GMM and Gabor features to detect the abnormalities effectively. In the Alex net for reasonable features information from F6and F7 layers were extracted. In DNN high end convolution layers are responsible for classification task, so information from F6 and F7 layers information were extracted and denoted as AF6 and AF7. 4096 features from Alex net with DNN and 64 dimensional vectors from GDP-FM (GM+GABOR+GLCM) were projected to JAYA algorithm for optimization. From these optimal features are selected and given to the Random Forest Classifier. This novel method E-RFC approached to maximum accuracy of 94.8% on an average with limited features and a small number of labeled data when compared to the other machine learning algorithms available in the literature.

Fig. (iv) Graphical Representation of Classifier performance with different features

In this method Public dataset DIARET 1B used for training and testing data set. Priya et al used geometrical based ROI extraction and ANN with SVM classifier and performed on single fundus with accuracy of 89.60%. With GLCM approach the authors Lachoure et al classified with accuracy of 82% on fundus images. Romany et al, performed two level classification task with Alex Net based DNN with SVM and achieved an accuracy of 91.03 and 94.40 % with feature optimization with Linear Discriminant Analysis (LDA) and Principal Component experimented on Kaggle data set and achieved an accuracy of 91 and 94% respectively. From the Fig.(iv) it can be observed that our work outperforms with limited number (400) of images with predefined dataset. The prominent and robust features are selected from segmenting ROI using combination of GLCM, GMM and Gabor features. Our work achieved maximum accuracy of 94.8% accuracy with optimized features (analogous to dimensionality reduction ) with Random Forest Classifier shown in Fig.(v).

Fig. (v) Graphical Representation of Classifier performance

CONCLUSION:

This paper demonstrates an enhanced frame work with a high accuracy to diagnose Diabetic Retinopathy with disease severity using ANN based Deep Neural Networks by detecting abnormalities in Diabetic Retinopathy with Fundus photographs. This work was carried out on publicly available data sets with their ground truth images in MATLAB 2015b. In future data augmentation can be applied to observe scale invariance. Current research on medical field mainly has a great focus on the usage of machine learning algorithms to assist clinicians to diagnose the disease at an early stage with minimum time span, so that preventive methods can be applied to reduce the disease severity. In future this method may be applied to multi class eye diseases related to diabetes like Glaucoma and age related macular edema and also need to focus for multiple disease classification related to Diabetic retinopathy.

REFERENCES:

1. Shaik Naseera, G.K. Rajini, B. Venkateswarlu, Jasmin Pemeena Priyadarisini M. A Review on Image Processing Applications in Medical Field. Research J. Pharm. and Tech. 2017; 10(10): 3556-3560, DOI:10.5958/0974-360X.2017.00644.8

2. S. K. Mittal, P. Nishant, A. Agrawal, S. Kumari, P. Kumar, and A. Chawhan. Community screening for diabetic retinopathy in uttarakhand, india, through targeted camps–a retrospective survey. Indian Journal of Community Ophthalmology. 2020; 1: 19–21.

3. Anannya Bose, Susanta Paul, Dibya Das, Tathagata Roy, Vinay Kumar Pandey. Recent Advancement of Nanomedicine for Diabetic Retinopathy: A Review. Research Journal of Pharmacy and Technology. 2023; 16(7): 3507-0.

4. Ankita Gupta, Rita Chhikara. Diabetic Retinopathy: Present and Past. Procedia Computer Science. 2018; 132: 1432-1440 https://doi.org/10.1016/j.procs.2018.05.074

5. Nandhini Murali, Abinaya S.K, Saveetha. V. Knowledge, attitude and practice of Diabetic Retinopathy among type II diabetic patients of South Indian population. Research J. Pharm. and Tech. 2017; 10(9): 3017-3021

6. A. H. Vyas and V. Khanduja. A Survey on Automated Eye Disease Detection using Computer Vision Based Techniques. 2021 IEEE Pune Section International Conference (PuneCon). 2021; 1-6. doi: 10.1109/PuneCon52575.2021.96864

79.

7. P. Vijay Daniel, D. Pamela, P. Kingston Stanley, J. Samson Issac. Digital Diagnosis of Diabetic Retinopathy using Fundus Images. Research J. Pharm. and Tech. 2019; 12(2): 717-722.

8. Reza AM. Realization of the contrast limited adaptive histogram equalization (CLAHE) for real-time image enhancement. Journal of VLSI Signal Processing Systems for Signal, Image and Video Technology. 2004; 38(1): 35–44

9. García-Lamont F, Cervantes J, López-Chau A, Ruiz S. editors. Contrast Enhancement of RGB Color Images by Histogram Equalization of Color Vectors’ Intensities. International Conference on Intelligent Computing. Springer, 2018

10. M. Kim and M. G. Chung. Recursively separated and weighted histogram equalization for brightness preservation and contrast enhancement. IEEE Transactions on Consumer Electronics. 2008; 54(3): 1389-1397. doi: 10.1109/TCE.2008.4637632.

11. Kuldeep Singh, Rajiv Kapoor, Sanjeev Kr. Sinha. Enhancement of low exposure images via recursive histogram equalization algorithms. Optik. 2015; 126(20): 2619-2625.

12. Ganesan P, B.S. Sathish, L.M.I. Leo Joseph, K.M. Subramanian, V. Kalist5, K. Vasanth. Retinal Blood Vessels and Optical Disc Segmentation in Branch Retinal Vein Occluded Fundus Images Using Digital Image Processing Techniques. Research J. Pharm. and Tech. 2019; 12(4): 1901-1906.

13. R. G. Bozomitu, A. Păsărică, V. Cehan, R. G. Lupu, C. Rotariu and E. Coca, Implementation of eye-tracking system based on circular Hough transform algorithm, 2015 E-Health and Bioengineering Conference (EHB), 2015; 1-4. doi: 10.1109/EHB.2015.7391384.

14. Christopher Jose, D. Aju. A Hybrid Method for Diabetic Retinopathy Diagnosis through Blood Vessel Extraction and Exudates Identification from 2D Fundus Image’. Research J. Pharm. and Tech. 2018; 11(3): 1147-1152.

15. H. Wan, H. Wang, B. Scotney and J. Liu. A Novel Gaussian Mixture Model for Classification. 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC), 2019; 3298-3303, doi: 10.1109/SMC.2019.8914215

16. A. Sopharak, B. Uyyanonvara, S. Barman, T.H. Williamson. Automatic detection of diabetic retinopathy exudates from non-dilated retinal images using mathematical morphology methods. Comput. Med. Imag. Graph. 2008; 32(8): 720-727.

17. J. Kaur, D. Mittal. A generalized method for the segmentation of exudates from pathological retinal fundus images. Biocybernetics and Biomedical Engineering. 2018; 38(1): 27-53.

18. M. Iyapparaja, P. Sivakumar. Detecting Diabetic Retinopathy exudates in digital image processing Hybrid Methodology. Research J. Pharm. and Tech. 2019; 12(1): 57-61.

19. A. D. Fleming, S. Philip, K. A. Goatman, J. A. Olson and P. F. Sharp. Automated microaneurysm detection using local contrast normalization and local vessel detection. IEEE Transactions on Medical Imaging. 2006; 25(9): 1223-1232.

20. Bhimavarapu U, Battineni G. Automatic Microaneurysms Detection for Early Diagnosis of Diabetic Retinopathy Using Improved Discrete Particle Swarm Optimization. J Pers Med. 2022 12(2):317. doi: 10.3390/jpm12020317. PMID: 35207805; PMCID: PMC8878235.

21. Veena Mayya, Sowmya Kamath S․, Uma Kulkarni. Automated microneurysms detection for early diagnosis of diabetic retinopathy: A Comprehensive review.Computer Methods and Programs in Biomedicine Update. 2021; 1: 2666-9900, doi.org/10.1016/j.cmpbup.2021.100013.

22. B.D. Venkatramana Reddy, T. Jayachandra Prasad. Color-Texture Image Segmentation Algorithms based on Hypercomplex Gabor Analysis. Research J. Engineering and Tech. 2011; 2(2): 77-86.

23. Haralick, R. M., Shanmugam, K. and Dinstein, I. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics. 1973; 3: 610–621, doi:10.1109/TSMC.1973.4309314

24. R. Roslan and N. Jamil. Texture feature extraction using 2-D Gabor Filters. 2012 International Symposium on Computer Applications and Industrial Electronics (ISCAIE), 2012; 173-178. doi: 10.1109/ISCAIE.2012.6482091.

25. Shaukat, N.; Amin, J.; Sharif, M.I.; Sharif, M.I.; Kadry, S.; Sevcik, L. Classification and Segmentation of Diabetic Retinopathy: A Systemic Review. Appl. Sci. 2023; 13: 3108. https://doi.org/10.3390/app13053108

26. Mujeeb Rahman, K.K.; Nasor, M.; Imran, A. Automatic Screening of Diabetic Retinopathy Using Fundus Images and Machine Learning Algorithms. Diagnostics. 2022; 12: 2262.

https://doi.org/10.3390/diagnostics12092262

27. F. Fernández-Navarro, M. Carbonero-Ruz, D. Becerra Alonso and M. Torres-Jiménez. Global Sensitivity Estimates for Neural Network Classifiers. IEEE Transactions on Neural Networks and Learning Systems. 2017; 28(11): 2592-2604.

28. Uddin, S., Khan, A., Hossain, M. E. and Moni, M. A. Comparing different supervised machine learning algorithms for disease prediction. BMC Med. Inform. Decis. Mak. 2019;1: 1–16

29. G-K, R. and Lavanya, R. Diabetic Retinopathy Image Classification Using Transfer Learning. Lecture Notes in Electrical Engineering. 2021; 700: 2511–2524. doi: 10.1007/978-981-15-8221-9_234

30. Khalifa NEM, Loey M, Taha MHN, Mohamed HNET. Deep Transfer Learning Models for Medical Diabetic Retinopathy Detection. Acta Inform Med. 2019; 27(5): 327-332. doi: 10.5455/aim.2019.27.327-332. PMID: 32210500; PMCID: PMC7085308.

31. F. Fernández-Navarro, M. Carbonero-Ruz, D. Becerra Alonso and M. Torres-Jiménez. Global Sensitivity Estimates for Neural Network Classifiers. IEEE Transactions on Neural Networks and Learning Systems. 2017; 28(11): 2592-2604.

32. Yaqoob, M.K.; Ali, S.F.; Bilal, M.; Hanif, M.S.; Al-Saggaf, U.M. ResNet Based Deep Features and Random Forest Classifier for Diabetic Retinopathy Detection. Sensors. 2021; 21: 3883.

33. Yaqoob, M. Kashif, Syed Farooq Ali, Irfan Kareem, and Muhammad MoazamFraz. Feature-based optimized deep residual network architecture for diabetic retinopathy detection. In 2020 IEEE 23rd International Multitopic Conference (INMIC)-2020; 1–6. IEEE, 2020

34. Himansu Das, Bighnaraj Naik, H.S. Behera, A Jaya. Algorithm based wrapper method for optimal feature selection in supervised classification. Journal of King Saud University - Computer and Information Sciences. 2022; 34(6): 3851-3863.

Received on 19.10.2023 Modified on 18.03.2024

Research J. Pharm. and Tech 2024; 17(9):4443-4448.

DOI: 10.52711/0974-360X.2024.00686