Deploying Deep Learning Models Using Serverless Computing for Diabetic Retinopathy Detection

Abstract

The incidence of diabetes is increasing at an alarming rate across the world. As a result, cases of diabetic retinopathy (DR) are on the rise, a complication of diabetes that in its most severe form can lead to blindness. The lack of specialized labor for diagnosis, essential for the successful treatment of the disease, brings the need to study alternatives for diagnosis via computational means. Recent research on the use of Deep Learning for the detection of DR proves to be an important alternative to improve the use of specialized labor, by prioritizing the most serious cases. From this context, the work objective is to evaluate the performance and financial cost of alternatives based on serverless computing for the deployment of Deep Learning models for DR classification. Using the Amazon Sagemaker serverless inference service, optimizations and different configuration alternatives were considered, obtaining up to $9.4\%$ of financial cost reduction and up to $2.35{\times }$ performance boost. Finally, concepts such as containerization and infrastructure as code were used during the solution implementation, to allow the reproduction of deployment and experiments performed.