Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Secure Federated Learning Approaches to Diagnosing COVID-19

Rittika Adhikari, Christopher Settles

TL;DR

This work develops a HIPAA-compliant federated learning framework to diagnose COVID-19 from chest X-rays by training across multiple hospitals without sharing raw images. It employs a DenseNet-121-based classifier within a secure aggregation protocol, and uses Grad-CAM heatmaps to interpret model decisions in clinical terms. The study analyzes the impact of data distribution (IID vs non-IID) on performance and demonstrates that secure aggregation yields comparable accuracy to centralized baselines, while preserving patient privacy. The approach has practical implications for rapid, privacy-preserving collaboration across healthcare institutions in the diagnostic workflow.

Abstract

The recent pandemic has underscored the importance of accurately diagnosing COVID-19 in hospital settings. A major challenge in this regard is differentiating COVID-19 from other respiratory illnesses based on chest X-rays, compounded by the restrictions of HIPAA compliance which limit the comparison of patient X-rays. This paper introduces a HIPAA-compliant model to aid in the diagnosis of COVID-19, utilizing federated learning. Federated learning is a distributed machine learning approach that allows for algorithm training across multiple decentralized devices using local data samples, without the need for data sharing. Our model advances previous efforts in chest X-ray diagnostic models. We examined leading models from established competitions in this domain and developed our own models tailored to be effective with specific hospital data. Considering the model's operation in a federated learning context, we explored the potential impact of biased data updates on the model's performance. To enhance hospital understanding of the model's decision-making process and to verify that the model is not focusing on irrelevant features, we employed a visualization technique that highlights key features in chest X-rays indicative of a positive COVID-19 diagnosis.

Secure Federated Learning Approaches to Diagnosing COVID-19

TL;DR

This work develops a HIPAA-compliant federated learning framework to diagnose COVID-19 from chest X-rays by training across multiple hospitals without sharing raw images. It employs a DenseNet-121-based classifier within a secure aggregation protocol, and uses Grad-CAM heatmaps to interpret model decisions in clinical terms. The study analyzes the impact of data distribution (IID vs non-IID) on performance and demonstrates that secure aggregation yields comparable accuracy to centralized baselines, while preserving patient privacy. The approach has practical implications for rapid, privacy-preserving collaboration across healthcare institutions in the diagnostic workflow.

Abstract

The recent pandemic has underscored the importance of accurately diagnosing COVID-19 in hospital settings. A major challenge in this regard is differentiating COVID-19 from other respiratory illnesses based on chest X-rays, compounded by the restrictions of HIPAA compliance which limit the comparison of patient X-rays. This paper introduces a HIPAA-compliant model to aid in the diagnosis of COVID-19, utilizing federated learning. Federated learning is a distributed machine learning approach that allows for algorithm training across multiple decentralized devices using local data samples, without the need for data sharing. Our model advances previous efforts in chest X-ray diagnostic models. We examined leading models from established competitions in this domain and developed our own models tailored to be effective with specific hospital data. Considering the model's operation in a federated learning context, we explored the potential impact of biased data updates on the model's performance. To enhance hospital understanding of the model's decision-making process and to verify that the model is not focusing on irrelevant features, we employed a visualization technique that highlights key features in chest X-rays indicative of a positive COVID-19 diagnosis.
Paper Structure (28 sections, 3 figures, 3 tables)

This paper contains 28 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Medical Image Diagnostic Methods
  4. Secure Federated Learning Methods
  5. Federated Learning with Non-IID Data
  6. Federated Learning in Diagnostic Applications
  7. Methods
  8. Model Architecture
  9. Secure Federated Learning Infrastructure
  10. Secure Aggregation
  11. Attacks against ML models
  12. Experimental Setup
  13. Interpretability Check
  14. Sanity Check
  15. Effect of Data Distribution
  16. ...and 13 more sections

Figures (3)

  • Figure 1: This figure describes the ideal end-user pipeline of our final secure federated learning system. First, the patient feels symptoms of the illness and sets up an appointment at the doctor's office. Next, the patient goes to the radiologist and undergoes an initial screening. We then utilize model inference to diagnose the patient's X-Ray. Next, an expert verifies the correctness of the diagnosis by analyzing an interpretable heatmap of the patient's X-Ray provided by the model. Finally, the expert prescribes the appropriate medical action to begin to treat the patient's illness.
  • Figure 2: An illustration of the secure aggregation protocol. The coordinator machine sends the global model to each client, and each client responds with an encrypted version of the updated model. From the coordinator's perspective, the coordinator cannot decipher any individual model that is sent from a hospital, and can only learn the average of all the models by summing the models together to cancel out the encryption masks.
  • Figure 3: