Federated Learning for Accurate Labeling of Chest X-Ray Scans

Abstract

Federated learning is an increasingly common technique used within machine learning that allows multiple devices to collectively train a model without necessitating the centralization of data. This approach is highly valuable within medical tasks, where privacy concerns within patient datasets can be mitigated through the decentralization of machine learning training. Within past literature, there have remained difficulties in constructing well annotated, large chest X-ray datasets due to these patient privacy concerns. In this paper, we seek to demonstrate the validity of federated learning by training a deep learning model on decentralized Chest X-ray imaging data. We utilize the publicly available NIH Chest X-ray dataset to train our model. Five clients were trained over 10 rounds, and a ResNet-34 global model was initialized and moved to a GPU, where clients iterated over each round to update the model. We initialized a new parameter accumulation dictionary for each round that was outfitted with Secure Aggregation algorithm with in-built additive homomorphic encryption of local parameters towards parameter averaging. The model achieved a validation loss of 0.09 and an accuracy of 0.83. These results indicate that the outlined federated learning approach was able to approach benchmark clinical grade accuracy, demonstrating the effectiveness of federated learning in advanced medical imaging analysis with the preservation of patient privacy.

Keywords

• federated learning
• chest x-ray imaging
• medical imaging
• deep learning
• resnet-34
• patient privacy
• decentralized machine learning
• healthcare data
• distributed training
• validation accuracy