Authors - Madhuri Wakode, Geetanjali Kale Abstract - Deep learning has many applications in healthcare especially for disease prediction from complex images. One such application is to predict diseases from chest X-rays (CXR). These models need huge amounts of data available for training. A single healthcare facility may struggle to collect sufficient data to build robust and efficient models. The apparent solution is collaboration among multiple healthcare facilities to use their data to build efficient models together. However, facilities may not want to share the patient’s sensitive data with other facilities or with the central server. Federated Learning (FL) allows multiple parties to build models without sharing their data with each other. FL allows parties to train the model locally on their private data and only share the trained model parameters to the server. Server averages the model parameters sent by all the parties to build a robust global model. Server sends this updated model to each party who then again trains the model locally. This process continues till the model convergence. We propose using federated learning average on a large CXR dataset for multi-label classification. Our results show that federated learning achieves the accuracy of ~82% as compared to ~91% of that of traditional centralized training method. FL with more robust algorithms and larger datasets, can achieve performance comparable to centralized approach with an added advantage of collaborative learning with privacy preservation.
