EdgeAI: Diabetic Retinopathy Detection in Intel Architecture

Diabetic retinopathy is a leading cause of blindness among working-age adults. Millions of people suffer from Diabetic Retinopathy in India. More dreaded situation is faced by the population in rural India where access to quality healthcare is limited. AI comes to the rescue in those situations where initial diagnosis can be performed without much manual intervention. Early detection of this condition is critical for good prognosis. In this paper we propose a solution using UP2 board (Edge device based on x86 architecture) where AI diagnosis can be performed on the local premise itself. We used PyTorch framework for training the model using EfficientNet-B4 network architecture. Trained model was optimized using Intel Distribution of Open-VINO, and hence there is no much compromise in execution time. Inference time for execution of single image in UP2 board is 0.2 sec. Our model achieved test metric performance comparable to baseline literature results, with sensitivity of 91.5% and specificity of 97.86%. For proof of concept we used open dataset from Kaggle 2019 competition hosted by Aravind Hospital, India.