In this research article, a brief insight into the detection of DR in human eyes using different types of preprocessing & segmentation techniques is being presented. There are a number of methods of segmenting the blood vessels that are present in the retina & once the retinal nerve fibers are segmented, one can detect whether the eyes are affected with diabetic retinopathy or not. In fact, this detection depends on the area of the RNFL network. If the total area of the nerve fibre is less, then it is affected with diabetic retinopathy (DR)& if the area of the nerve network is more, then the eyes are not affected with the diabetic retinopathy and hence it is normal. It is a well-known fact that diabetics assumes a vital job in the health of the human beings & affects each and every organ. One such organ in the human eye. This DR will give rise to vision loss in the human eye as the optic nerve is connected to the brain. The retinal fundus images are commonly used for detecting & analyzing of disease in disease affected images. Raw retinal fundus images are difficult to process by machine learning algorithm. Hence, a survey is being given here in this very context. This is a review paper / survey paper in which any researcher who reads this paper, he / she can get some idea about the disease in the human eye, how it gets affected, symptoms, etc... In fact, to say, the paper can be thought of as an introductory paper about the diabetic retinopathy& its background. Various research analyzers have chipped away at this theme of the topic till now. To start with, 100’s of research papers were collected from various sources, studied @ length & breadth and a brief review of the eye disease issues was being made & presented here in a nutshell. In the sense, the recent works done by various authors across the globe is being presented here in this context so that this review article serves as the base for any researcher who is working in the field of ophthalmology could define their own new research problem. One of the important organs of the human being is the eye. It has to be noted that if the eyes are not there, then the whole world would be dark & the human life even though it is existing will be a waste. Different types of the diseases occur in the eyes. One of the deadliest diseases which occurs in the eyes is the DR. This disease is the second largest disease which is occurring amongst the human beings as per the WHO – United Nations survey. Hence, utmost importance has to be given to the eye care. This disease occurs due the reduction of the nerve area in the retina. If the area of the RNFL decreases, then the optic nerve which is connecting to the brain gets damage, leading to the loss of vision. In this paper, a mere introduction is given to the diabetic retinopathy disease. Hence, an exhaustive review is given w.r.t. the said disease, which is the topic of research taken by us as a part of the Ph.D. program.
What it does
It takes image of eye through glass lens, and convert into a retino graph. After this we analyze the vein structure, and blood clustering inside the eye ball to predict the present stage of diabetic patient.
How we built it
The color retina images are downloaded from the Kaggle website. The training dataset contains 35126 high resolution images under a variety of imaging conditions. These retina images were obtained from a group of subjects, and for each subject two images were obtained for left and right eyes, respectively. The labels were provided by clinicians who rated the presence of diabetic retinopathy in each image by a scale of “0, 1, 2, 3, 4”, which represent “no DR”, “mild”, “moderate”, “severe”, “proliferative DR” respectively. As mentioned in the description of the dataset, the images in the dataset come from different models and types of camera, which can affect the visual appearance of left vs. right. Also, the dataset doesn’t have the equal distributions among the 5scales. As one can expect, normal data with label “0” is the biggest class in the whole dataset, while “proliferative DR” data is the smallest class.
We trained our convolutional neural network in Table 1 on a single Tesla-P100 GPU. For nonlinearity, we use leaky (0.01) rectifier units following each convolutional layer. The networks are trained with Nesterov momentum with fixed schedule over 250 epochs. For the nets on 256 and 128-pixel images, we stop training after 200 epochs. L2 weight decay with factor 0.0005 are applied to all layers. As we treat the problem as a regression problem, the loss function is mean squared error. The convolutional networks have untied biases. Batch size is fixed at 32 for all networks. 4. Following the evaluation setting , the quadratic weighted Kappa score is adopted as the performance metric of prediction. Specifically, the predicted regression values are discretized at the thresholds (0:5; 1:5; 2:5; 3:5) to obtain integer levels for computing the Kappa scores and making submissions. All the features mentioned in Section 4.2 were also adopted in our model training.
Challenges we ran into
Getting good quality images from the sensor
Accomplishments that we're proud of
Very good accuracy of more than 80%
What we learned
Pytorch and Deep learning
What's next for Diabetes Detection using Retinopathy using Pytorch
Building the camera to collect images and send for test data