Overview / Usage

Early detection and diagnosis of DR requires the detection of the severity level of DR but that is a challenging task for both the medical experts as well as for computer-aided diagnosis systems since extensive domain expert knowledge is required for detection of the severity levels of DR. Efforts to reduce its prevalence and more effectively manage its health consequences are further undermined by the fact that roughly half of all people with diabetes are currently undiagnosed. However, early detection of DR through periodic screening is the key to maintaining optimal health and quality of life as well as reducing costs due to timely treatment. Scarcity of resources and trained professionals makes the screening for diabetic retinopathy a major challenge for the public health system. Hence, automated screening systems may have to be considered for periodic screening for Diabetic Retinopathy. India, in particular, has a large rural population, battling illiteracy and lack of awareness about the disease. Lab facilities are generally poor and treatment is expensive. This provides us a rationale for the development of a classification system for diabetic retinopathy images that could be deployed in community based screening in medically under-served communities

Methodology / Approach

Our solution is based on DNN and is running using Intel’s Neural Compute Stick. The Intel’s Neural Compute Stick comes with a “neural network accelerator” on the chip, which allows it to tackle deep learning applications based on neural networks. It is capable of delivering 1 TFLOPS of performance in a small package designed to fit into hand-held devices. The Neural Compute Stick, with a pre-trained CNN model downloaded over the USB interface, functions as an inference engine that can be used for providing inferences on data presented to it. This will provide a powerful DR classification tool when presented with new fundus camera images. It can function as a stand-alone, edge computing device, where connectivity is not available or not reliable. Since any pre-trained CNN model can be downloaded to the NCS based Inference Engine, it can be used in a variety of other image classification tasks. Some of the potential application areas are: Macular Edema, Cervical Cancer, Oral Cancer etc.,

Technologies Used

Deep Learning Frameworks. Intel distribution of Python, Intel distribution of TensorFlow, Keras, NCS-1 / NCS-2, OpenVINO, NCS-SDK

Other links

• https://www.calligotech.com