Deploying Machine Learning Inference on Diabetic Retinopathy in Binary and Multi-class Classification

Diabetic Retinopathy is a condition of the person's eye that causes vision loss and blindness in diabetic people. Among adults between 20–74 years, this disease most of the time acts as the cause of blindness. Also, it is the leading reason behind cecity among working-aged adults worldwide. An efficient health care system contributes to an essential part of the country's economy, development, and industrialization. It affects one out of three persons with diabetes. The longer a person has diabetes, the higher their risk of developing some ocular problem. The mobile and web application built adds assistance to the people for endorsing their qualms and knowing about their condition's severity. Since most of the users can be from rural areas, we built applications that work both offline and online. The inference is to select a configuration based on a single input. To decrease the magnitude of the mobile application model, we implemented MobileN et architecture. It is considered a light weightedmodel as it has fewer parameters than others, making it suitable for the application objective. The accuracy of MobileNet is 76 percent. ResNet 50 architecture is used in the web application, which requires an internet facility to predict the output. Because of the increase in the parameters, it provides higher accuracy when compared to the MobileN et model. Kaggle eyepacs dataset is given as input for both models. The dataset is divided into five classes: No Diabetic Retinopathy, mild, moderate, severe, and proliferative. The models classify the images given as input and display one of these class labels as a result. An oversampling technique has been implemented to overcome the problem of data bias. NodeJS and python have been used to build web applications and android studio for mobile applications. The time taken to display the result in both applications is less than 10 seconds. The size of the mobile application has been reduced from 120MB to 25 MB. This application is helpful in rural areas wherever they lack facilities to run tests while obtaining the results on their phones and any web browser by simply up- loading the scanned image of the patient's eye.