Medical Image Classification using Federated Learning

Abstract

Medical data requires strict privacy protection, making it difficult to make large medical im age datasets. With less samples to train a model, many deep learning models fails to classify the data accurately. Many Deep Learning models need to collect all the data at a data center, but it has privacy concerns. So Federated Learning (FL) is helpful in such situations to train a model when there is a shortage in the dataset. Federated learning allows clients to indepen dently build a local deep neural network (DNN) model using local data before working to gether to combine a global DNN model at the centralized server. Federated Learning works on the concept of Differential Privacy. Differential Privacy Helps to collect and share the aggregate information of the users,by maintaining the user privacy. As Federated Learning is decentralized learning and data is distributed across many clients in an uneven manner. In this paper we have built a model which classifies the medical image data using federated learning which takes images of brain tumors and skin lesions. The accuracy for brain tumors is 95%andthatforskin lesions is 88%. Weshowthatourapproachmaintainspatient privacy while simultaneously providing the highest level of accuracy.