Efficient Local Cloud-Based Solution for Diabetic Retinopathy Detection

1. Introduction

The diabetes problem associated with retina is Diabetic Retinopathy(NIH National Eye Institute, 2014). It can be avoided by detecting damage of retina (leaking fluid or blood). Loss of vision can be avoided by early detection of diabetic retinopathy.

It progress through four stages:

• i.

  Mild nonproliferative retinopathy: A little amount of swelling near blood vessels of theretina which are called micro-aneurysms. This mainly occurs at a very early stage of the disease. Some amount of fluid leakage may also take place into the retina.

• ii.

  Moderate nonproliferative retinopathy: Increase in the swelling and distortion of theblood vessels due to their inability to transport blood. The retina’s appearance can change and may lead to Diabetic Macular Edema.

• iii.

  Severe nonproliferative retinopathy: In this case, some blood vessels are completelyblocked thereby eliminating the blood supply to the areas near the retina. In some cases this may lead to growth of newer blood vessels near the retina.

• iv.

  Proliferative diabetic retinopathy (PDR): The most advanced stage of Diabetic Retinopathy is PDR where the proliferation of recently-grown blood vessels occurs. Some blood vessels also start to grow inside the surface of the retina, eventually turning into a gel-like patch, possibly filling the eyeball and causing blindness. 

Convolutional Neural Network (CNN) is a branch of deep learning and has a massiveapplication in the field of image analysis and medical imaging for detection of disease. They are neural network architectures specifically designed for handling data with some spatial topology (e.g. images, sound, videos, character sequences in text, 3D voxel data etc). Convolutional Neural Network also has a collection of hidden layers. Specialized low power microscopic, indirect ophthalmoscopes capture fundus images (Kanski, J.J., Bowling, B., 2011), yielding high quality images in terms of resolution, clarity and contain a high pixel value. Fundus image is used in proposed work.

Figure 1.

Normal Retina VS Diabetic retinopathy

It is observed that there is a need of an automated detection system which can easily detect the presence or an absence of the disease and can also save time and effort that is wasted in the manual detection of diabetic retinopathy. Also the use of CNN (Andonová M., et al.,2017) in the detection of the disease by using the retinal images makes it an important approach towards the early detection of diabetic retinopathy(AndonováM.,et al.,.2017). The design of this system intentionally automates the entire diagnosis process to classify the level of diabetic retinopathy through use of the patient’s fundus images.The ratio of people affected with the disease to the number of eye specialist who can screen these patients is very huge. Hence there is the need for an automated diagnostic system to measure diabetic retinopathic changes in the eye, so the affected personcan be referred to the specialist for further intervention and treatment,It has been observed that the whole process is manual and takes a lot of time to diagnose the disease so the initialObservations from ophthalmic practice reveals that this manual practice for diagnosis is time consuming. Also, using Convolutional Neural Network enables the large dataset of various classes of DR to serve as a foundation for further input of a real time dataset to diagnose specific patients’ eye conditions and risk for blindness. As the dataset of the system receives more data, resulting in an oversampling of the image dataset, researchers can rebalance the population of the dataset, to increase the accuracy of diagnosis suggested by this local cloud-based solution for diabetic retinopathy detection. Eventually the accuracy of 96% is also a source of motivation of extending the modules to the next stage where other neural network models can also be included for better accuracy.