Abstract:Deep learning (DL) has immense potential to revolutionize healthcare. Several Federated DL solutions have been proposed to access massive repositories of private data, without transferring subject data from the host devices. There remain concerns about potential privacy violation via "reconstructing" individually-identifiable subject data by exploiting model weights from one host. We propose a methodology for federated DL that addresses this risk through cloud-server architecture design. The study is a fully federated replication of a previous DM classifier derived on 300 head-and-neck cancer patients. Only CTs (1 slice from mid-tumor) from a The Cancer Imaging Archive public collection was used. Two hospital cohorts were used for training, and 2 other cohorts were used as validation. A "privacy-by-architecture" federated DL prototype was developed by extending an open-source distributed learning system VANTAGE6. An authenticated researcher submitted a learning task to a master server, that distributed a DL network to be trained to data hosts. Host devices were Ubuntu machines with 1 GPU each and data stored in its local XNAT repository. Training and validation were randomly split across both data hosts. The DL model consisted of 3 fully-connected convolutional blocks and 1 sigmoid output layer. In each iteration of training, synchronous averaging and update of model weights were managed through a separate aggregation server connected to each data host via a key and encrypted web interfaces. The aggregation server prevented direct connections between hosts, researcher and master server while training, so that none could peek at interim model weights generated by others. When finished, final globally-averaged weights were stored on the master server for retrieval by the researcher, but aggregator contents were erased. We demonstrated functional equivalence of a federated DL classifier to its centralized version, without transferring any subject-level data from data host machines. Further, by preventing access to interim model weights, we added an additional layer of privacy protection which we call "privacy-by-architecture". The centralized learning and federated learning models delivered statistically equivalent performance in terms of accuracy, loss and area under receiver-operator curve. 5-fold cross validation mean AUC/accuracy of centralized and federated DL models are 0.88 vs 0.85 and 0.88 vs 0.86 respectively. We demonstrated a privacy-preserving federated DL methodology that enhances privacy protection by preventing visibility of model weights during training; globally-averaged model weights reduce the risk of privacy violation by " reconstruction" attacks on the data hosts. Our architecture may be further enhanced by adding differential privacy and encryption in future.

Bilateral Proxy Federated Domain Generalization for Privacy-preserving Medical Image Diagnosis

A Survey on Domain Generalization for Medical Image Analysis

Universal Medical Imaging Model for Domain Generalization with Data Privacy

Privacy-Preserving Constrained Domain Generalization via Gradient Alignment

Federated Ophthalmic Disease Diagnosis Based on Weighted Averaging and Local Optimization

Improving the Classification Effect of Clinical Images of Diseases for Multi-Source Privacy Protection

MH-pFLGB: Model Heterogeneous personalized Federated Learning via Global Bypass for Medical Image Analysis

Feasibility of Privacy-Preserving Federated Deep Learning on Medical Images

GRACE: A Generalized and Personalized Federated Learning Method for Medical Imaging

Medical Federated Model with Mixture of Personalized and Sharing Components

Federated Distillation for Medical Image Classification: Towards Trustworthy Computer-Aided Diagnosis

Medical Federated Model with Mixture of Personalized and Shared Components

Federated adversarial domain generalization network: A novel machinery fault diagnosis method with data privacy

MLA-BIN: Model-level Attention and Batch-instance Style Normalization for Domain Generalization of Federated Learning on Medical Image Segmentation

Model-level Attention and Batch-instance Style Normalization for Federated Learning on Medical Image Segmentation

Frequency-Based Federated Domain Generalization for Polyp Segmentation

FedDP: Privacy-preserving method based on federated learning for histopathology image segmentation

Client-Level Differential Privacy via Adaptive Intermediary in Federated Medical Imaging

Medical report generation based on multimodal federated learning

Asymmetrical Reciprocity-based Federated Learning for Resolving Disparities in Medical Diagnosis

Federated Domain Generalization with Generalization Adjustment