Description:Field of Invention
In digital image processing, an input picture or video file is processed to produce a final product that may be another image, a collection of attributes or characteristics associated with the original image. In most cases, it is a reference to processing images digitally. Both contemporary ophthalmology and diabetes rely heavily on computer vision and image processing methods. Retina, a light-sensitive tissue located at the back of the eye, can have its blood vessels damaged by diabetic retinopathy (DR), an eye condition. Among those under the age of 70, it is a major cause of vision loss and deterioration, and it is a prevalent complication of diabetes. The primary cause of diabetic retinopathy (DR) is an irrational rise in blood glucose levels, which damages the endothelium lining the blood vessels and makes them more permeable to the retina. Retinal detachment occurs as a result of DR expansion. Less effective treatment options will be available to DR patients until visual impairment starts, as they will not be aware of any symptoms. In order to treat and maybe prevent vision loss caused by DR, laser photocoagulation is used in its early phases.
Background of the Invention
Advancements in computer-aided technology are steadily rising in the field of medical imaging diagnosis systems. The use of image processing and other forms of modern computing has found particular use in ophthalmic screening. Human anatomy includes the five senses of sight, hearing, taste, and smell as well as touch. Every sense is critical for our daily lives, but vision in particular is essential since it allows us to see our physical surroundings. "Vision" is the universal term for sight.The retina receives the focused light, which activates the rods and cones. After then, the optic nerve receives these impulses from the retina. The visual images are processed by the brain through impulses transmitted by the optic nerve. A thorough understanding of the eye's structure is necessary for diagnosing and treating eye diseases. The receptor organ, resembling a camera, is housed in the spherical human eye. There are three layers that cover the eyeball. The cornea and sclera make up the fibrous tunic, the outermost layer.
The ability to see, which entails the brain's conversion of light impulses into images, is the most important sense for humans. The disease known as diabetic retinopathy damages the blood vessels in the retina, a light-sensitive tissue located in the back of the eye. Diabetic eye disease is a prevalent complication of the disease and a major contributor to visual impairment and blindness in adults younger than 70 years of age.
The primary cause of diabetic retinopathy (DR) is an irrational rise in blood glucose levels, which damages the endothelium lining the blood vessels and makes them more permeable to the retina. Retinal detachment occurs as a result of DR expansion. Less effective treatment options will be available to DR patients until visual impairment starts, as they will not be aware of any symptoms. In order to treat DR and maybe avoid vision loss, laser photocoagulation is used in its early phases.
Automated vascular segmentation in the retina has been a hot topic for the last ten years. Segmentation methods have been developed by Fraz et al. (2012) and can be classified as either supervised or unsupervised. While segmentation is taking place, supervised algorithms are already aware of the ground truth in a training set of images, but unsupervised algorithms learn on the go. The unsupervised method makes use of algorithms including matched filtering, vessel tracking, morphological transformations, contour model, Laplacian operator, and perceptual transformation approach, among others mentioned in related publications. Training the network with datasets obtained from public databases and hospitals, labeled as normal, mild, moderate, and severe, is the foundation of the proposed study, which is based on supervised learning methods.
In order to facilitate research into early DR detection, Yen (2008) created an intuitive interface. This study aids ophthalmologists by following the protocol of the Early Treatment Diabetic Retinopathy Study (ETDRS). Based on their expertise and background, various graders were given the fundus photos to review. They are able to sort more efficiently and throughput more effectively. This sorting system was built using the suggested hybrid intelligent system. The characteristics of the spot lesion group are the only ones used in the first study. Its viability and performance are later proven when it is combined with the ETDRS.
When it comes to retinal pictures, Bhaskar and Kumar (2015) created a classifier called FLANN that can identify hard exudates. Luminosity Contrast Normalization performed the pre-processing. We compare the proposed classifier's experimental results to those of Radial Basis Function and Multi-Layer Perceptron. This indicates that the proposed FLANN classifier outperformed the others when it came to recognizing exudates.
Summary of the Invention
Those who have had diabetes for longer than ten years are more likely to develop diabetic retinopathy, a condition that damages the eyes. Delays in therapy, misunderstandings, and other complications may result from the time-consuming process of early detection of DR, even for clinicians who are well-prepared.The goal of the suggested approach is to use different segmentation and machine learning algorithms to determine the severity of early-stage diabetic retinopathy. Using the green channel and CLAHE, the color retinal fundus image is prepared for analysis. Feature extraction is the next step after image preprocessing. Combining all the lesions and performing the grading process follows preprocessing. Features such as blood vessels, microaneurysms, hemorrhages, and exudates are extracted. Some conditions have had the DR severity graded. To determine if diabetic retinopathy is mild, moderate, severe, or normal, classifiers like support vector machines and logistic regression are trained.
Brief Description of Drawings
Figure 1: Block Diagram of Proposed Methodology.
Figure 2. Work flow of preprocessing
Detailed Description of the Invention
To avoid diabetic retinopathy, a condition that affects the eyes, patients with diabetes should have their eyes checked often. Avoiding blindness due to diabetic retinopathy is possible with prompt treatment. Since DR does not have a major impact on the eye in its early stages, it can be treated. Blindness can occur as a result of the course of DR disease. Therefore, it is crucial to identify and treat diabetic patients early on to avoid visual loss. Patients with DR should have their eyes checked regularly. It takes a lot of time and money to screen each patient personally. Automated detection of glucose retinopathy early symptoms including blood vessel, microaneurysms (MAs), hemorrhages (HAs), and exudates is achieved by state-of-the-art medical image processing algorithms. Immediate and effective therapy can be achieved by the diagnosis of these anomalies, potentially averting blindness in the eye. The majority of the studies relied on fundus color retinal images or FA images captured during fundoscopy with dilated pupils. In order to overcome the difficulty of detecting small MAs and thin arteries in low contrast pictures, the current research applies high contrast enhancement. The suggested classifiers are equipped with the features necessary to detect and grade abnormalities; these features are derived from independent procedures for microaneurysm, hemorrhage, exudate, and blood vessel recognition.
Figure 1 shows that CLAHE and the green channel are used to preprocess the color retinal fundus image. Feature extraction is the next step after image preprocessing. The features such blood vessels, microaneurysms, hemorrhages, and exudates are extracted after preprocessing. Then, all the lesions are combined and graded. Some conditions have had the DR severity graded. Normal, mild, moderate, and severe diabetic retinopathy severity levels are determined by training classifiers like SVM and MLP.
Preprocessing is the initial stage of the given task. Because of abnormalities like camera settings and noise, the original retinal image could have undesirable distortions, low lighting, and noise. Depending on these pictures might change the diagnosis and the final verdict. The development of new arteries and extremely tiny lesions are sometimes undetectable to retinal specialists in certain areas. Near the optic disc, the retina's brightest area, the intensities may change. Experts may fail to detect certain abnormalities and newly formed blood vessels on the optic disc's outer surface due to the inadequate lighting. The retinal fundus images undergo preprocessing to eliminate unwanted artifacts such as noise and poor lighting. Improving the image's quality and features requires preprocessing. Resizing images, extracting green channels, applying median filters, and CLAHE are all part of it. The procedure for the preprocessing method is illustrated in Figure 2.
The multi-dimensional photographs come from a variety of public sources and hospitals. In order to prepare them for additional processing, these photos are standardly sized at 128×128 pixels. In this case, size is determined empirically and then normalized. Retinal pictures are processed by excluding the red and blue parts of the image in order to increase contrast. During the preprocessing step, the green channel is utilized to highlight the most prominent areas and achieve the highest level of contrast between the optic disc and retinal tissue. The values of the green pixels in the input image are therefore extracted and stored in a matrix format. Retinal images become more susceptible to noise and other aberrations when local datasets are used for diagnosis. Therefore, it is believed that removing distortion and noise is a crucial step before making a diagnosis. The outcome might be skewed as a result. This case calls for the median filter. It is a non-linear filter that keeps an image's crisp edges while reducing noise and other distortions. Using a tiny sliding window, the median filter sorts nearby pixels by intensity and uses their median value as the new value for the central pixel. It is also good at preserving the original retinal image while eliminating horizontal artifacts, salt-and-pepper noise, and outliers.
Retinal photographs taken during an examination sometimes have issues with contrast, lighting, and other photographic abnormalities. The fundus retinal pictures show a greater amount of light in the middle than in the periphery or sides.There is a lot of light in the middle of the areas, but very little light around the edges and corners. As a result, the retinal image's poor contrast enhancement may cause some lesions to go unnoticed since the center portion of the image and the OD are overemphasized. To improve the image's contrast throughout the whole retina, Contrast Limited Adaptive Histogram Equalization (CLAHE) is used. The contrast enhancement approach known as contrast limited adaptive histogram equalization (CLAHE) divides the entire region into several smaller portions of uniform size. It then improves the contrast in each of these smaller regions. By applying histogram equalization to each region, we change the distribution of gray values and bring out the image's hidden features. Consequently, this method improves the picture quality. The application of CLAHE to the green channel enhances the visibility of the fundus image's finer details, including information on blood vessels, microaneurysms, hemorrhages, and exudates.
Morphology, the study of forms, is an integral aspect of image processing. Typically applied for tasks such as noise reduction, edge identification, picture enhancement, and segmentation, mathematical morphology primarily relies on operators employed in binary image analysis. Combining the binary picture and the structural element with any of the set operations—intersection, union, inclusion, complement—is what morphological operations are all about. Processing and encoding of the input image in the structuring element is dependent on form attributes. By comparing the correct pixel value from the input image with its neighbors' pixel values, the pixel value of the output image is computed. Erosive, dilatant, opening, and closure are the four main categories of morphological procedures.
The pictures may still have blood vessel, MA, HA, and EX anomalies after the preprocessing procedures. These features are first extracted using a variety of segmentation approaches, and only after that are they permitted to be used for classification. Preprocessing, morphological operation, thresholding of vessels, background removal, and contour identification are all steps that fundus photographs pulled from databases must go through before blood vessel segmentation can begin.
Following the extraction of blood vessels, microaneurysms, hemorrhage, and exudate, a support vector machine (SVM) classifier is employed to differentiate between good and bad images. With this classifier, over-fitting and empirical risk are less likely to occur. With this outcome as a foundation, you can achieve a remarkable performance. After extracting areas of blood vessels, microaneurysms, hemorrhages, and exudates from segmented color fundus pictures, they are ready to be processed with the support vector machine classifier. The goal of this tool is to find the normal, mild, moderate, and severe stages of diabetic retinopathy problems based on the training data. We combine the features used for segmenting regular and abnormal pictures and save them in a matrix. The support vector machine classifier uses this matrix as input and returns a testable version of its training output in the same format. After the test photo properties are extracted and combined, a matrix is created for testing reasons. The severity of the disease is trained into the SVM classifier before the test image is input. In order to differentiate between typical and non-typical images, a support vector machine classifier employs the kernel function. In terms of the anomalous image output, there is a scale that goes from mildly affected to severely affected.
A fully connected neural network, or MLP, is one in which every node in one layer is linked to every other node in the network. It has at least three levels: input, output, and a hidden layer or layers that pass input from one layer to another. Every layer save the input layer uses a non-linear activation function for its nodes. Numerous hidden units, or perceptrons, make up each hidden layer. The perceptrons form an array in the single hidden layer. The goal of developing MLP was to provide a perceptron structure to classes with more than two variables. For NN-based classifications, a collection of learning rules is laid out. Each node in the network communicates with the others. Every node takes in a larger quantity of data, processes it using simple mathematical operations, and then outputs a single value. Each node is given a weight, and the output is determined by adding up all of the inputs that are weighted. These weights are learned and used in the recognition process during training. The backpropagation process is executed using the threshold function. The whole feedforward NN's error function is found by the back propagation procedure. After determining the output and hidden units' error terms, it uses these terms, along with the learning rate, to fine-tune the weights. Iterative training on the training set is continued until the error value drops to a minimum. , Claims:The scope of the invention is defined by the following claims:

Claim:
The System/Method for Detection and Grading of Diabetic Retinopathy using Hybrid Machine Learning Techniques comprising the steps of:
a) A method for detection of small MAs and detection of thin vessels in low contrast images.
b) A method for Grading the severity of DR.
c) A method for Object identification and calculation of time required for detection.
2. According to claim 1, the System/Method for Detection and Grading of Diabetic Retinopathy using Hybrid Machine Learning Techniques as claimed in claim1, led to the design of high contrast enhancement is used
3. As per claim 1, the System/Method for Detection and Grading of Diabetic Retinopathy using Hybrid Machine Learning Techniques as claimed in claim1, color retinal fundus image are preprocessd using green channel and CLAHE.
4. According to claim 1, the Design of System/Method for Detection and Grading of Diabetic Retinopathy using Hybrid Machine Learning Techniques as claimed in claim1, hybrid machine Learning classifiers SVM and MLP are designed.