An Empirical Study on Privacy-Preserving Swarm Learning for Cataract Detection

Cataract is one of the most prevalent eye diseases in today’s day and age. Roughly 51% of the world’s cases of blindness is a result of cataract even though it can be prevented if detected at an earlier stage. This paper proposes a privacy-preserving cataract detection model with a maximum accuracy of 98%, which predicts if a patient has cataract by processing their retinal scan using a decentralized framework called Swarm Learning. The main focus of this paper lies in preserving the privacy of patient data while creating an accurate and robust model. The performance of a machine learning model directly correlates with the amount of training data. Healthcare data is highly distributed in different hospitals, and data privacy laws make it difficult to collect data and train a model in a centralized setup. Hence this work focuses on a decentralized architecture, where each hospital acts as an individual node with its own training data and a VGG-19 cataract detection model. The model proposed in this paper provides promising results even though it contains non-independent and identically distributed data, i.e. the amount of data and the nature of data varies with each node.