deewan07/diabetic-retinopathy-detection-using-deep-learning

Diabetic-Retinopathy-Detection-Using-Deep-Learning

Diabetic Retinopathy is a leading disease-causing vision-loss globally. Retinopathy has several stages if not rehabilitated in time it could be severe. It originates from the diabetic disease, an increase of sugar level in the blood vessels of the retina. This paper is all about the smart proposed model which is trained to detect the stage of Diabetic Retinopathy in patients. simply by putting the fundus images into our model, it will not only detect Retinopathy disease but also tell which stage it is. Generally, DR can be separated into two major stages: non-proliferative DR (NPDR) and proliferative DR (PDR). Our model is capable to detect 5 stages of DR namely No DR, Mild DR, Moderate DR, Severe DR, and Proliferative DR. It will be very beneficial for doctors as well as for researchers to detect Diabetic Retinopathy patients at an instant. Generally, it is time-consuming for doctors to analyze each and every tiny nerve cell from fundus images.

The proposed model uses Densenet Convolutional Neural Network to train the model. First, the dataset is n preprocessed. The images of the dataset are converted from BGR(blue, green, red) to RGB(red, green, blue) format. Now The dataset is split into train and test dataset 85 percent and 15 percent respectively. The images are resized to 224*224 size for the Densenet model. Next, we create MULTI LABELS for the images .To create multi labels, we use logical OR between adjacent images. The image which is diagnosed with Severe Diabetic Retinopathy (id = 3) will have its multi labels as [1 1 1 1 0]. In order to further improve the dataset quantity , we augment the dataset by Flipping the images horizontally . Flipping the images vertically . Fill mode = Constant. Random zooming of the images by 0.15.

Conclusion

The paper presented a deep learning approach to detect diabetic retinopathy. The system utilizes a DenseNet powered training model, which can be utilized for quick identification of Diabetic Retinopathy. The system at first identifies the prominent features of the input retina image. On the basis of the features , the system predicts the type of retinopathy based upon its training. The results indicate that the proposed system can successfully detect the type of retinopathy with accuracy between 96-98(in percentage) based on different stages of the disease. Further, the initial results of disease prediction also shows a good agreement with actual results given by clinicians. The future work includes to get or create a better dataset for class 3 and 4 of the model. Overall the system will help in detecting the type of retinopathy and based on the results, the patients can be given proper medication based on the seriousness of the disease.