Federated Learning for Colorectal Cancer Prediction

The availability of datasets pertaining to various fields has increased significantly in the past decade, but there still exists a problem in getting datasets pertaining to the medical field as most of the data needs to be confidential and there exists laws which ensure a patient's data privacy. Federated learning (FL) proves to solve this problem via a client-server architecture by enabling distributed training of clients, without any data exposure. In this paper, we apply the FedAvg (FederatedAveraging) [1] algorithm on the PathMNISTv2 [2] dataset for predicting colorectal cancer. We also present a refined convolutional neural network (CNN) architecture for accurate predictions on the PathMNISTv2 dataset. We have studied the effects on IID (Independent and Identically Distributed) and Non-IID (Non-Identically Independently Distributed) distributions in a distributed environment. We have also compared these results with a centralized model and demonstrate that FedAvg achieves similar results in a distributed setting. We anticipate our study to enable additional healthcare studies driven by vast and diverse data, and illustrate the efficacy of FL at such magnitude and task complexity as a paradigm shift for multi-site partnerships, eliminating the need for data sharing.