828
Retinal Based Pathology Analysis Using Deep Learning Approaches
Nivetha Na, Aarthi Mb, Devisri Gc , Evangilin Sd, and Gowri Se
a,b
Assistant Professor, Department of Computer Science And Engineering K.Ramakrishnan College of Technology,
c,d,e UG Student, Department of Computer Science And Engineering, K.Ramakrishnan College of Technology
Article History: Received: 10 January 2021; Revised: 12 February 2021; Accepted: 27 March 2021; Published online: 16 April
2021
_____________________________________________________________________________________________________ Abstract: Deep learning is an excellent approach to biomedical image segmentation tasks that help to identify and quantify
patterns. Segmentation of biomedical images typically involves partitioning an image into multiple regions representing anatomical objects of interest. Biomedical image processing is a very broad field. Here, input is given as an image to extract features from an image to detect diseases. Unsupervised learning technique is used because it enables the model to uncover trends and facts that were previously undetected on its own. Human eye retina can provide valuable information about human health. The condition of the retinal vessels has been found to be a good indicator of overall health. For the prediction of retinal disease, it is classified as a retinal background, retinal vessels and optic disc. The disease can be predicted by this classification using the Singular Spectrum Analysis, Deep Neural Network, KNN Classifier and Eclipse Fitting algorithm. Singular Spectrum analysis was carried out to predict cross-section profiles for the detection of micro aneurysms. Diabetic diseases are classified under the KNN classification. Deep Neural Network algorithm is used for predicting cardiovascular disease and other disease such as glaucoma, stroke. Using the Eclipse Fitting algorithm, the optical disc and cup pixels are segmented to check whether or not a person's health is normal.
Keywords: Deep Learning, Image Processing, diabetic disease, cardio vascular disease, glaucoma.
1. Introduction
Mathematical operations and signal processing forms are used to perform image processing in image science. An picture, image series, or film (such as a photo or video clip) serves as the input, and the output is a set of images or special characteristics. photo Picture settings Typically, the picture is viewed as a three-dimensional signal with a three-dimensional time (or z-axis). Digital image processing is the most available, but optical and analogue image processing are also options. This article addresses several different approaches that can be done for any of these methods. Computer science for disease identification and diagnosis is becoming increasingly popular in the field of life sciences. Detecting the disease took too long in the past, and it was ineffective. The principle of machine learning (ML), which has been used in recent biomedical studies, has been greatly enhanced by increasing precision and minimizing time. Machine learning can also be used to automatically diagnose illnesses in scanned images. We hope to diagnose retinal disorders using retinal imaging as a part of this study. A deep learning (DL)-based approach for detecting retinal disease in retinal images is presented in this paper. Retinal imaging is used to detect and classify diabetic retinopathy (DR), age-related macular degeneration (AMD), macular degeneration, retinoblastoma, retinal detachment, and retinitis pigmentosa. Automated diagnosis of retinal disease may be a significant step forward in terms of early detection and disease progression avoidance. Several sophisticated methods for segmentation and automated identification of road signs and retinal lesions have been established in the past. Advances of machine learning and computational medical imaging in the area of ophthalmology, on the other hand, have opened up a whole new avenue of study for researchers. Retina imaging can also reflect a scientific advance in eye care. The retina, blood vessels, and optic nerve behind the eye can be digitally imaged by the optometrist. This aids in the diagnosis and prevention of diseases that can damage the eyes and general health. Glaucoma, macular degeneration, diabetes, and elevated blood pressure are only a few examples. The most subtle changes in the anatomy under the eye are normally visible through retinal imaging techniques. Individuals are often put in danger by vascular disease, which often poses a public health risk. The need for sophisticated medical imaging stems from a need to learn more about and treat these disorders. The retina, blind spot, macula and fovea, and posterior pole are all situated in front of the lens on the eye's inner surface. Microcirculation is the only part of the body that can be seen up close and personal. In science, ophthalmoscopes and fundus photographs are often used. It can be used to locate blood vessels in the fundus while developing a new unsupervised fuzzy algorithm for blood vessel tracking. It employs linguistic concepts (such as "vascular" and "non-vascular") to automatically track fundus arteries and solve recorded angioplasty and contour analysis problems. The most important tool for determining ship and non-ship areas along a ship's contours is the fuzzy C-means clustering algorithm, which provides entirely preprocessed results. A protocol for
evaluating the viability of the identified container, processing accessories, and divisions is also included. A semi-automatic technique to assess the geometric and step characteristics of the continuous vascular tree is then seen on a clinical illustration of the fundus. The size is generated using binary images collected using previously specified segmentation techniques. Split tree skeletons can be modified and generated, branches and intersections can be recognized, tree segments can be labeled, and sequence code can be saved. The user picks a tree to assess and decides if it is an artery or vein. Then measure the height, area, and angle of a single tree segment using an automatic process. Geometric records and descriptions of the relationships between the retinal vessels' continuous Branches Are Established.
2. Relatedwork
According to Fan Huang et al. [1.], Diabetic retinopathy (DR) is that the most frequent explanation for blindness in jobs within the u. s. and also the world organization. Diabetic retinopathy should be diagnosed early (through scans, for example) and treated appropriately to avoid vision loss and blindness. in step with US forecasts for the following decade, trained ophthalmologists won't be able to have a mean age and also the number of diabetics in each cohort won't be fully serviced, a minimum of within the short term. the employment of digital cameras rather than outpatient care by ophthalmologists, especially the utilization of computers to read these images for early diagnosis, has often been contentious. the quality of an ophthalmologist's test is critical because it'd eventually assess the existence of DR within the retina likewise as other pathological features like glaucoma and cataracts. Some people are aware of digital imaging and will be reading ophthalmologist images, but they will not be at home with computer algorithms. Computer algorithms, on the opposite hand, tend to be a minimum of as successful as human readers, and are expected to satisfy the wants of virtually half diabetics who don't even have daily health exams. Since the dataset lacks access to verity state of the disorder, the accuracy of all algorithms can only be attributed to human measurements. I can only interpret human reading, which is wrong from the standpoint I've said, but it can function a reference dataset. there's no other thanks to get a deeper understanding of this reference dataset's current medical state. to attenuate the possible effect of coaching data on results, all algorithms use the identical training data. The DR detection algorithm can attain efficiency and parallel sophistication reminiscent of an accomplished single retinal reader, but with little detectable additional cognitive enhancements.
According to Jaydeep De et al. [2], diabetic retinopathy (DR) is the main cause of blindness in developing countries and is one of the diabetic problems of most long-term disease patients. While early diagnosis and constant monitoring of diabetes are critical, appropriate therapies for DR can be used. The diagnosis of DR is made by examining the retinal photograph (fundus) Manually classifying these pictures and assigning DR intensity is time-consuming and resource-intensive. The appearance of a microaneurysm (MA) in the retina is the disease's first and most common symptom. The retinal MA acts as a tiny circular red mark. This paper reflects on the automated diagnosis of retinal Alzheimer's disease and proposes a mechanism for accomplishing this goal. This approach has proved to be fully competitive with most cutting-edge technology, according to the findings of a free online competition. The ability to compete in the DR classification process, understanding Alzheimer's disease is crucial. This is how Alzheimer's disease is used to determine whether or not a picture of a patient's eye is stable. As a result, there is a lot of literature on computer-aided diagnosis (CAD) approaches for diagnosing DR and other eye-related diseases, and retinal imaging science is a very successful and surprising area in the digital field. Not at all. Group of image processors. MA has a maximum diameter that is scientifically defined and is usually less than the diameter of the major optic nerve vein. Small circular spots resembling Alzheimer's disease in size and appearance may arise from the junction of small blood vessels. Vessel fragments may be cut from the vessel tree and shown as tiny dark objects in a variety of shapes. Almost all cutting-edge approaches have image pre-processing steps such as noise reduction, filtering, and shading correction. A typical approach for minimising noise is to use Gaussian masks for convolution and median filtering. Although the proposed approach does not entail any preprocessing, it has been discovered that considering image smoothing before the actual detection stage is useful.
Erik J. Bekkers et al. [3], investigated the identification of microaneurysms in optical colour fundus images. This is a vital first step toward immediately covering a common diabetes complication, diabetic retinopathy (DR). Several strategies for making these detections have been published in the past, but the same data cannot be balanced. The ROC's aim is to conduct a series of competitions focused on some of the main challenges of automated identification of retinal diseases. The first competition focuses on the identification of
830
microaneurysm, to expand on the previous paragraph's definition. The highest probability of lesions (irrelevant) in an image in our studies reflects the likelihood that the whole image will contain DR signals. Automated mask (DR) analytical tasks can be used to detect microaneurysms in the optics, and automated masking is a common issue in diabetes. Many methods for performing this identification have been published in the past, but none of them are incompatible for equivalent results. The findings of a major international microaneurysm identification contest conducted as part of the Retinopathy Online Challenge (ROC), a multi-year online contest covering all facets of DR detection, were discussed in this article.Jaydeep De and a coworker [4 ], developed the microaneurysms (MAs) in the retina is typically the first symptom of diabetic retinopathy, so early identification is crucial to stop blindness (DR). As a result, it's important to have automated Alzheimer's disease diagnosis in the screening software. Individual small circular particles with a diameter of 10-100 nm are commonly used to identify these lesions. In reality, they are represented by a large number of MA groups or fleets. MAS is normally composed of a group of agents that work in either a virtual or real-world environment. In dynamic and flexible organisations, agents interact with one another to organise operations. Self-adaptation is the product of a subject's adaptation to environment experiences and restrictions. Isolation and contact between agents create global consequences that influence other agents in the system. As a result, agents appear to converge on predictable, regular solutions that aren't completely known at the person level. Other functions that a single agent cannot have are enabled by the energy provided by the new phenomena. Using the MAS process, a new red lesion segmentation algorithm was proposed in this report. The findings of conventional algorithms can be obtained by detecting MA at the ends of blood vessels and examining the local interaction of agents. You may calculate standard FP counts and provide agent learning metrics for future algorithm creation by introducing a validation phase that uses the characteristics of this field. Using the MAS process, a new red lesion segmentation algorithm was proposed in this report. Traditional algorithms can achieve results primarily through the local association of agents by detecting MA at the ends of blood vessels. We suggested introducing validation steps to the domain definition to minimise the number of PFs and incorporating some agent learning methods to develop the algorithm further. To show the proposed method's theoretical effect, it must be compared to the ROC method. Our findings aren't perfect, but they can be seen as inspiration and compared to other studies.
Li Cheng et al. [5] issued this treatise that how to detect and describe blood vessels in fundus images using an automatic procedure. Ophthalmologists may use those resources to perform medical assessment, treatment review, and clinical testing. Our system varies from previous methods in that it separates the ship network by integrating the characteristics of ships from the area and around the globe. The Matched Filter Response (MFR) is coded in grayscale, with darker pixels indicating a stronger response. It is obvious that the central response of the non-vascular MFR picture is greater than the response of the left vessel. The MFR image is measured using the defined form, and thresholding is done using a new detection technique. The picture is cropped and certain area-based attributes are measured by the probe. If the probe detects a blood artery in the drill bit, it will segment and identify the constituent pixels at the same time. Our detection process, as opposed to the classifier-based method, allows for the measurement of pixels in various geographic configurations prior to final classification. The procedure identified isolates around two-thirds of blood vessels during retinal fundus imaging. In contrast to the previous approach, which depended entirely on global thresholds, the proposed method removes approximately half of the false positives while still lowering the true positive answer. The above is due to the small system for connecting the ships in order to break them apart. The suggested edge path tracking approach is used in cine angiography to segment the artery. To model edge paths, Markov chains are used. The SEL (Sequential Edge Linking) algorithm is used to find the best way across all feasible paths for a Markov model. Modify the model's functionality to represent path characteristics such as local curvature tolerance. The benefit of using this approach is that the grouping operation is effective with the real gradient value rather than the threshold response. As a consequence, no segmentation decisions are taken before a sufficient number of pixels can be classified. The drawback to this approach is that the branches aren't modelled, so each branch must be tracked and analysed separately.
S.Rahmath Nisha and S.Kiruthiga [14] presented computer based method to detect cancer and its stages. Stages of cancer was detected by analyzing various images using deep convolution technique . The same method can be used to detect images of retina.
3. Existing Methodologies
The diabetic retina plays a vital role. Nonproliferative diabetic retinopathy is another name for diabetic retinopathy. Furthermore, for diabetics, robotic or computer-assisted retinal diagnosis assists ophthalmologists in screening vast numbers of patients. While there are many patients, the workload on local ophthalmologists is minimal. As a result, automated screening should be implemented to minimise disease incidence and enable
ophthalmologists to successfully diagnose and treat the disease. Building such an integrated device necessitates the creation of many modules that examine anatomical features of the retina such as the fovea centralis, blind spots, blood vessels, and common diabetic disorders such as bleeding and microaneurysms. Furthermore, early automated identification of microaneurysms may ease ophthalmologist discomfort. To define the exudate, select a function and use the SimpleBay and Support Vector Machine (SVM) classifiers. First, a training collection of 15 features was derived from samples of positive and negative bleed pixels and used to train the naive Bayes model. It then extracts a special SVM from a single feature set of the naive Bayes classifier and proceeds to apply the omitted features to the classifier. Find the best hyperparameter combination using grid inspection, such as the training error tolerance for each feature combination and the width of the radial base functions.
3.1 DISADVANTAGE
• There is a lot of computational complexity. • Hardware is used to diagnose the disease • The reliability of segmentation is lower. • Fault tolerance is a possibility.
4. Proposed Methodologies
The macula, blind spots, and blood vessels are the retina’s most prominent anatomical features. The eye’s macula is the tiniest portion of the retina with the clearest vision. Behind the eye, the macula of the retina is a layer of photosensitive tissue. Blood vessels are elastic channels or tubes that allow blood to circulate across the body. Eye diseases such as glaucoma and diabetic retinopathy can be diagnosed by analysing the blood vessels within the eye. Vascular networks are usually drawn by hand, but this is a time-consuming process that requires skill and expertise. Through automating the procedure, you will ensure accuracy and, most importantly, save time spent on manual scanning by skilled technicians and doctors. Implements an automated technique for analysing blood vessels in retinal images to diagnose cardiovascular disease. It employs the active contours concept to reduce noise, enhance image quality, track the sides of blood vessels, assess their surroundings, and detect heart disease. To segment blood vessels and measure boundaries, use a neural network model. This is a strong and accurate infinite boundary active contour model with mixed domain terms for ship segmentation that can produce excellent results. It can be an effective method for measuring blood vessels and treating a wide range of vascular diseases. The CRAE and CRVE constants of retinal vessel opening were investigated. By estimating the mean retinal vascular opening correlates with the existence or absence of categorical variables or a rise in the quartile of continuous variables, analysis of covariance is used to predict blind spot (OD) cardiovascular disease. retina is a word that is used to describe the As the ganglion cell’s axons defocus and form the optic nerve, this happens. The optic nerve is responsible for transmitting visual information from image receptors to the brain. On the retina’s surface, it’s a slight blind spot of around 3 mm on the nasal side of the macula. Just a small portion of the eye is unaffected by light. It is a symptom of a variety of different eye diseases. Actually, there are a number of eye disorders. Diabetic retinopathy and glaucoma are two disorders that cause vision loss. These diseases are also being researched by ophthalmologists. Vision loss may be minimised if these disorders are detected and handled early. This approach would need more ophthalmologists as the population rises and ages. To detect these diseases at an early stage, an automatic blind spot recognition system is required. Blind spot detection is crucial in assessing the seriousness of diseases like glaucoma and diabetic retinopathy. The most well-known characteristics of retinal fundus imaging should be considered when assessing blind spots. The blind spot is divided into two sections: a bright central area (known as the cup) and a peripheral area (called the neural retina). The cup region is critical for glaucoma diagnosis because it helps assess the relationship between the cup and the intervertebral disc. Alternative methods for predicting glaucoma have been suggested. The detection limit for blind spots and the splitting of cups are two commonly used approaches. Linear discriminant analysis, contrast-limited adaptive histogram equalisation, green channel extraction, and an ellipse fitting process are all used in the proposed blind spot boundary recognition system. The ellipse fitting method includes feature extraction. To improve the image, adaptive histogram equalisation with contrast limitation is used. Finally, use an ellipse operation to detect blind spot boundaries in the resulting image. The green channel of the input image is used for cup segmentation. Using the watershed conversion for cup splits. Determine the cup to disc ratio after assessing the blind spot and the cup boundary. Cup-to-disc ratios usually vary from 0.1 to 0.4. When the cup-to-disc ratio is greater than 0.4, it suggests a problem, such as glaucoma. Methods that are accurate yield results without the need for user interaction.
832
4.1 IMAGE ACQUISITIONThis module's aim is to capture digital images. Stroke, diabetes, arteriosclerosis, cardiovascular disease, and hypertension are all disorders that can be detected and diagnosed using human retinal photos. People with vascular disease are often in life-threatening conditions, and society is dealing with severe public health issues. The retina requires a lot of imaging, and detecting blood vessels is particularly critical. Vascular displacement (duration, distance, bifurcation pattern, etc.) not only provides accurate information about pathological changes, but also aids in the classification of disease severity and automatic diagnosis. I'm going to do it. A retinal picture must be uploaded. The retina, blind spot, macula and fovea, and posterior pole make up the lower portion of the inner focal point within the eye in front of the lens. An ophthalmoscope or a fundus image may be used to examine the eye. Multiple layers of neurons connected by synapses make up the retina, which can be a complicated structure. The veins in the retina can be seen. Blood vessels in the early stages of cancer display abnormalities or deformities. The systemic arteriole and vein stenosis associated with elevated blood pressure levels is determined by the ratio of arteriole to vein diameter. He developed an image dataset for the proposed method's training and evaluation. He created this image dataset by combining publicly available datasets (including DRIVE and STAR). Each pixel is recorded using 24-bit colour at 760 x 570 pixels (standard RGB). It's easier to discern than a normal picture, for starters. Second, you must achieve some degree of success, and its unusual vascular presence can suggest clinical application. As you can see, a regular image consists of blood vessels, the optic nerve, the fovea, and the background, while an abnormal image consists of various objects of various shapes and colours caused by various diseases.
4.2 PREPROCESSING
Modify the picture in such a way that the chances of completing other processes are improved. To differentiate between black and white illumination, the grayscale conversion operation is used. Because coloured noise pixels and blurry pixels are the most common causes of colour retinal image noise, intermediate filters are commonly used to enhance and adjust vascular patterns used for retinal image preprocessing and vascular segmentation. Production, effectiveness improvement, and segmentation are all phases in the processing process. Images of the retina and blood vessel shapes. Because the edges and details of an image are particularly sensitive to human perception and are primarily composed of high frequency components, the visual quality of the image will be significantly reduced if the high frequencies are attenuated or completely removed. Emphasizing the high frequency components of an image, on the other hand, can improve visual efficiency. The technique of image sharpening emphasises the edges and details of a photograph. To improve regional contrast and sharpen images, image sharpening is commonly used in the printing and photographic industries. In theory, sharpening an image entails applying a logo to the first image that is proportional to the first image's high-pass filtered version. The original image is first filtered with a highpass filter to eliminate high-frequency components, and then a scaled version of the highpass filter's output is added to the original image to produce a clearer image than the original. I'm going to do it. It's worth noting that the signal's neutral field, that is, the signal's stable region, is unaffected.
4.3 SEGMENTATION
Computer vision methods such as active contour patterns (also known as snakes) are used to separate the contours of objects. Snakes are deformable splines in 2D pictures that can be noisy. The snake is deformed in this method to mitigate energy, the internal force resisting deformation, and the constraints and image forces. Push
the snake in the direction of the object's contour. Another two-dimensional technique is the active shape model, which uses a target distribution model to restrict the shape spectrum to a particular region learned from the training set. Finally, look for preprocessed retinal images in the split screen. Alzheimer's disease, as well as the most common interference targets (cross-vascular, retinal haemorrhage, small vessel fragments, and so on), may be a priority in Oman. To get these run candidates that display the actual shape and location of the object, you'll need to use an area expansion strategy. These measurements can, however, surpass nearby MAs and other properties, depending on your local threshold settings. We built a collection of statistical features from candidate profiles that can differentiate between real MA and non-MA to solve this problem. The proposed approach focuses on single spectrum analysis and considers the candidate object's principal axis when constructing the contour, overcoming these limitations significantly.
4.4 CLASSIFICATION
To reduce the effects of noise, the SSA configuration file also describes candidate object information. We discovered that the MA profile is more similar to the non-MA profile in all directions after discovering the SSA-based profile. To optimise the difference between MA and non-MA, the variance score is added to each section profile of the candidate object. It is impossible to examine improvements in veins and arteries without first distinguishing between them. To distinguish segmented vessels, a directional technique known as a convolutional neural network algorithm is used. Following the extraction of the blood vessels, functional carriers with properties that help the arteries and veins are created. In an image, you can extract features and label each centerline to represent arterial and venous pixels using the concept of centerline extraction. The ultimate goal is to assign one of the labels to arterial grade (A) and the other to venous grade (V) based on the results of the labelling process (V). Structural data is not only used to provide information about the ship's strength, but it can also be classified into A/V categories. This can be accomplished with the help of neural networks. The disc boundary and the cup form of the retinal image must also be determined.
4.5 DISEASE DIAGNOSIS
To diagnose disease, AVR reports are combined with CRAE and CRVE measurements. It is incorrect to equate the number of vessels in CRAE and CRVE with cardiovascular risk factors. Higher vital signs are the most important systemic determinants for small CRAEs, smoking is the most important systemic determinant for larger CRVEs, and higher vital signs, systemic inflammation, and obesity are actually the most important systemic determinants for larger CRVEs. It is a consideration. In a group of 80 healthy people, a recent study discovered a clear indirect relationship between renal function and retinal parameters (CRAE and CRVE). This means that target organ damage in the preclinical stage is a common determinant. This backs up previous research that found a connection between retinal vessel signs and unintentional hypertension. This indicates that a lower CRAE is related to an increased risk of unintentional hypertension. This is in line with previous research linking retinal vascular signs to accidental hypertension, in which a decrease in CRAE leads to clinical hypertension and subsequent signs of other organ damage. CRAE has numerous uses in other diseases, such as stroke and diabetes, in addition to predicting the importance of high blood pressure. A higher risk of stroke is related to systemic arterial stenosis, which is a measure of CRAE reduction. This module employs the KNN classifier to more accurately predict disease. The cutoff value can be used to predict whether a patient is diabetic or not. Type 1 diabetics and type 2 diabetics are two different forms of diabetes. Finally, use the right ratio to calculate the cup/disk ratio for glaucoma disease prediction.
4.6 ADVANTAGES
• Accurately determine disorders based on images of blood vessels in the retina. • Have better efficiency
• A simple method for detecting all diseases • All blood vessels are tracked in vessel monitoring.
5. Experimental Results
The suggested algorithm is evaluated in terms of saving storage and is applied in real-time settings. The findings can be shown using the .NET platform as the front end and SQL SERVER as the back end.
834
FIG 2: DISEASE PREDICTIONFIG 3: RETINAL DIABETIC PREDICTION
FIG 4: UPLOAD IMAGE
FIG 6: PREPROCESSING – MEDIAN FILTERING
FIG 7: NOISE REMOVED IMAGE
FIG 8: SPREAD SPECTRUM ANALYSIS
836
FIG 10: DISEASE PREDICTION – TYPE 2 DIABETICSFIG 11: VESSEL EXTRACTION
FIG 13: GRAPH POINTS
FIG 14: VESSEL CLASSIFICATION
838
FIG 16: OPTIC DISK SEGMENTATIONFIG 17: CUP SEGMENTATION
FIG 19: PATIENT REPORT DETAILS
FIG 20: HEART DISEASE REPORT
840
imaging. The proposed candidate filtering method will greatly reduce the number of non-MA candidates while still extracting more candidates at the vascular system's edge. To filter MA candidate configuration files, use the default SSA process. Implement ship segmentation post-processing steps. This move is used to locate the best forest and monitor all of the actual ships. Overcome intersection misdiagnosis by assessing the value of using structural details for A/V classification at the same time. Veins and arteries can be observed more closely in diseases and various systemic diseases, vein ratios can be measured, and these features can be analysed and forecast to study diseases and various systemic diseases. I'm going to do it. We also compared the preamble to this approach to other recently proposed methods and found that it produced better results. By performing SVM classification accuracy based on intensity characteristics, neural network-based approaches using ut show the value of using structural knowledge for A / V classification, with higher accuracy in retinal disease. Offers are made. In the future, we will expand this architecture to introduce various segmentation and classification algorithms to effectively communicate disease and integrate devices into the system.References
1. F. Huang, J. Zhang, E. J. Bekkers, B. Dashtbozorg, and B. M. ter Haar Romeny, “Stability analysis of fractal dimension in retinal vasculature,” in Proc. 2nd Int. Workshop Ophthalmic Med. Image Anal. (OMIA), Munich, Germany, Oct. 2015, pp. 1–8.
2. E. J. Bekkers, J. Zhang, R. Duits, and B. M. ter Haar Romeny, “Curvature based biomarkers for diabetic retinopathy via exponential curve fits in SE(2),” in Proc. 2nd Int. Workshop Ophthalmic Med. Image Anal. (OMIA), Munich, Germany, Oct. 2015, pp. 113–120.
3. J. De, H. Li, and L. Cheng, “Tracing retinal vessel trees by transductive inference,” BMC Bioinf., vol. 15, no. 1, p. 20, 2014
4. J. De, T. Ma, H. Li, M. Dash, and C. Li, Automated Tracing of Retinal Blood Vessels Using Graphical Models (Lecture Notes in Computer Science). Berlin, Germany: Springer, 2013, pp. 277–289.
5. J. De et al., “A graph-theoretical approach for tracing filamentary structures in neuronal and retinal images,” IEEE Trans. Med. Imag., vol. 35, no. 1, pp. 257–259, Jan. 2016.
6. E. Türetken, F. Benmansour, and P. Fua, “Automated reconstruction of tree structures using path classifiers and mixed integer programming,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2012, pp. 566–573.
7. B. Dashtbozorg, A. M. Mendonca, and A. Campilho, “An automatic graph-based approach for
artery/vein classification in retinal images,” IEEE Trans. Image Process., vol. 23, no. 3, pp. 1073–1083, Mar. 2014.
8. G. Azzopardi, N. Strisciuglio, M. Vento, and N. Petkov, “Trainable COSFIRE filters for vessel delineation with application to retinal images,” Med. Image Anal., vol. 19, no. 1, pp. 46–57, 2015. 9. L. Cheng, J. De, X. Zhang, F. Lin, and H. Li, “Tracing retinal blood vessels by matrix-forest theorem of
directed graphs,” in Medical Image Computing and Computer-Assisted Intervention (Lecture Notes in Computer Science). Cham, Switzerland: Springer, 2014, pp. 626–633.
10. Kalitzeos, G. Y. Lip, and R. Heitmar, “Retinal vessel tortuosity measures and their applications,” Experim. Eye Res., vol. 106, pp. 40–46, Jan. 2013.
11. Aarthi, M. and Bhuvaneshwaran, A., 2021. Iot Based Drainage and Waste Management Monitoring and Alert System for Smart City. Annals of the Romanian Society for Cell Biology, pp.6641-6651.
12. Pavithra, M., Sindhana, A.M., Subajanaki, T. and Mahalakshmi, S., 2021. Effective Heart Disease Prediction Systems Using Data Mining Techniques. Annals of the Romanian Society for Cell Biology, pp.6566-6571.
13. Tresa, M., Francina, S., Jerlin Oviya, V. and Lavanya, K., 2021. A Study on Internet of Things: Overview, Automation, Wireless Technology, Robotics. Annals of the Romanian Society for Cell Biology, pp.6546-6555.
14. S.Kiruthika,S.Rahmath Nisha ,2020, Automated Oral Cancer Detection and Classification using very Deep Convolutional Neural Network Algorithm.. TEST Engineering and Management 83 (March-April 2020), pp.20019 – 20027
