Source-linked AI summary
Dynamic Fusion based Federated Learning for COVID-19 Detection
Weishan Zhang, Tao Zhou, Qinghua Lu, Xiao Wang, Chunsheng Zhu, Haoyun Sun, Zhipeng Wang, Sin Kit Lo, Fei-Yue Wang
TL;DR
Privacy restrictions limit shared medical-image datasets, while default federated learning can impose high communication costs and struggle with heterogeneous client data. The paper proposes dynamic fusion-based federated learning that selects clients by local performance and training time, and reports better accuracy, robustness, fault tolerance, and communication efficiency than default federated learning.
Problem
Privacy concerns restrict sharing diagnostic images across medical institutions, limiting training datasets, while default federated learning can have high communication costs and reduced performance under client data heterogeneity.
Method
The approach combines a federated-learning architecture with dynamic fusion that selects clients by local model performance and schedules aggregation using participating clients’ training time.
Results
The proposed approach performs better than default federated learning in model accuracy, fault tolerance, robustness, and communication efficiency.
Takeaways & Limitations
Dynamic fusion is reported as a feasible way to improve COVID-19 detection federated learning while preserving data privacy and reducing communication overhead through less model uploading.
Abstract
from arXiv · showhide
Medical diagnostic image analysis (e.g., CT scan or X-Ray) using machine learning is an efficient and accurate way to detect COVID-19 infections. However, sharing diagnostic images across medical institutions is usually not allowed due to the concern of patients' privacy. This causes the issue of insufficient datasets for training the image classification model. Federated learning is an emerging privacy-preserving machine learning paradigm that produces an unbiased global model based on the received updates of local models trained by clients without exchanging clients' local data. Nevertheless, the default setting of federated learning introduces huge communication cost of transferring model updates and can hardly ensure model performance when data heterogeneity of clients heavily exists. To improve communication efficiency and model performance, in this paper, we propose a novel dynamic fusion-based federated learning approach for medical diagnostic image analysis to detect COVID-19 infections. First, we design an architecture for dynamic fusion-based federated learning systems to analyse medical diagnostic images. Further, we present a dynamic fusion method to dynamically decide the participating clients according to their local model performance and schedule the model fusion-based on participating clients' training time. In addition, we summarise a category of medical diagnostic image datasets for COVID-19 detection, which can be used by the machine learning community for image analysis. The evaluation results show that the proposed approach is feasible and performs better than the default setting of federated learning in terms of model performance, communication efficiency and fault tolerance.
I. INTRODUCTION
The paper addresses privacy-limited medical-image datasets and default federated learning’s communication and heterogeneity challenges for COVID-19 detection. It proposes dynamic fusion-based federated learning to improve accuracy, communication efficiency, and fault tolerance.
- Motivation: Privacy concerns prevent medical institutions from sharing diagnostic images, limiting dataset size for COVID-19 model training.The motivation focuses on CT and X-Ray analysis, whose performance depends heavily on training-dataset size and diversity.
- Problem: Default federated learning avoids exchanging raw data but can incur substantial model-update communication costs and underperform with heterogeneous client data.Clients train locally and send updates for central aggregation, while local data remain at each institution.
- Approach: The paper proposes a dynamic fusion approach that selects clients using local model performance and schedules fusion using participating clients’ training time.Clients upload only when their current local model outperforms its previous version, while the server configures waiting time from recent average training time.
- Contributions: The paper also summarizes medical diagnostic image datasets for COVID-19 detection for use by the machine learning community.
- Results: The proposed approach achieves better detection accuracy, fault tolerance, and communication efficiency than default federated learning.
II. DYNAMIC FUSION-BASED FEDERATED LEARNING FOR COVID-19 DETECTION
This section introduces a dynamic fusion-based federated learning approach for CT-image COVID-19 diagnosis. It covers the system architecture and a dynamic model-fusion method for client participation and aggregation scheduling.
- Scope: The approach targets CT scan image analysis for diagnosing COVID-19 infections.
- Design: The architecture describes component interactions, while dynamic model fusion selects participating clients and schedules aggregation based on training time.
- Organization: The section divides the approach into architecture and dynamic fusion components.
A. Architecture
The architecture coordinates a central server and clients that train models on locally stored diagnostic images. Local performance assessment and server-side scheduling determine which updates enter each aggregation round.
- Components: The system contains a central server and clients, with the server coordinating federated learning and clients training local models using local data and resources.
- Data and Jobs: Clients collect and preprocess diagnostic images locally, while the server creates training jobs containing model code, aggregation count, and initial waiting time.
- Client Assessment: A client uploads its model only when the current local model performs better than the previous version.
- Aggregation: Clients that miss the waiting-time deadline are excluded from the round, after which the server aggregates accepted updates and returns the global model.
B. Dynamic Fusion
Dynamic fusion uses client-side performance decisions and server-side waiting-time selection to reduce unnecessary or late model uploads. The waiting time is based on participating clients’ recent training times.
- Client Participation: Clients decide whether to join an aggregation round according to the performance of their newly trained models.
- Client Selection: The central server excludes clients that do not upload their model updates within the configured waiting time.
- Scheduling: From the second round, each client’s timer is based on the average training time of participating clients in the previous round.
- Aggregation: The server proceeds with aggregation without clients that fail to complete training within the configured time.
III. EVALUATION
The evaluation uses categorized CT and X-ray COVID-19 image datasets and distributes selected images across three clients for federated experiments.
- Datasets: Table I summarizes a category of medical diagnostic image datasets for COVID-19 detection.The datasets are used for quantitative evaluation of the proposed approach.
- Experiment environment: Table II presents the experiment environment used for the federated evaluation.
- Datasets: The evaluation datasets contain 746 CT images and 2,960 X-ray images for COVID-19 detection.The CT collection includes 349 positive and 397 negative cases; the larger X-ray collection includes COVID-19, negative, and viral pneumonia images.
- Experiment environment: The experiment uses one central server and three clients with different configurations.The collected images are divided into training and test sets, with client dataset sizes of 600, 900, and 1,300 images.
A. Accuracy
Across experiments with three model architectures, dynamic fusion-based federated learning generally achieves higher accuracy than default federated learning and remains robust under interference.
- Experimental setup: The experiments compare GhostNet, ResNet50, and ResNet101 under DF FL and D FL.Six dataset groups produce 18 experiments across the three model architectures.
- Accuracy: 14 of 18 experiments achieve higher accuracy with DF FL than with D FL.DF FL has lower accuracy in four groups by 1.711%, 0.57%, 0.57%, and 1.141%, respectively.
- Fault tolerance: DF FL maintains relatively steady results and higher accuracy than D FL when negative cases are marked as positive COVID-19.The interference is introduced in the fourth dataset group for each model.
- Accuracy: DF FL achieves higher accuracy than D FL in 14 experiments on randomly cropped test sets.In the remaining groups, DF FL is lower by 0.57%, 1.331%, 0.951%, and 1.141%, respectively.
B. Training Time
Dynamic fusion reduces training time mainly for larger models with more parameters, while GhostNet shows no training-time reduction because its model transmission is already short.
- B. Training Time: ResNet50 training time is reduced by 8-10 minutes, while ResNet101 training time decreases by 25-30 minutes.The experiments used 90 client training epochs and a maximum model upload/download speed of 10 MB/s.
- B. Training Time: GhostNet shows no reduction in training time because its smaller parameter count yields an average model transmission time of 2.2s.In contrast, larger ResNet models require more time to transmit model updates.
- B. Training Time: The proposed approach significantly reduces training time when the network is poor and the model has large amounts of parameters.
C. Communication Efficiency
Dynamic fusion reduces communication overhead by uploading fewer model updates, with the largest savings for parameter-heavy models.
- C. Communication Efficiency: For GhostNet, dynamic fusion decreases uploads by an average of 61 and reduces upload time by 110-160s, reaching 1/3 of default federated learning time.
- C. Communication Efficiency: For ResNet50, dynamic fusion decreases uploads by an average of 80 and reduces upload time by 900-1200s, reaching 1/10 of default federated learning time.
- C. Communication Efficiency: For ResNet101, dynamic fusion decreases uploads by an average of 78 and reduces upload time by 3200-4200s, reaching 1/16 of default federated learning time.
- C. Communication Efficiency: Dynamic fusion reduces communication overhead through less model uploading, with stronger effects for complicated models containing more parameters.
IV. RELATED WORK
Prior work established federated learning for cross-device, cross-silo, and medical-image applications, but communication efficiency and model accuracy remained insufficiently addressed for COVID-19 detection.
- IV. RELATED WORK: Federated learning originated in cross-device applications and later expanded to cross-silo learning across organisations and data centers.
- IV. RELATED WORK: Research has sought to reduce federated learning communication rounds through selective, scheduled, asynchronous, temporally weighted, and controlled aggregation methods.
- IV. RELATED WORK: Studies have addressed non-IID data through training-data clustering, multi-stage local training, and multi-task learning, while other work designs client incentives.
- IV. RELATED WORK: Federated learning has been applied to CT and X-Ray COVID-19 detection, but prior studies did not consider communication efficiency and model accuracy together.
V. CONCLUSION
The proposed dynamic fusion-based federated learning approach improves COVID-19 detection while preserving data privacy and performs better than default federated learning across several evaluation dimensions.
- V. CONCLUSION: The approach improves model accuracy and communication efficiency while preserving data privacy for COVID-19 detection.
- V. CONCLUSION: Evaluation indicates better performance than default federated learning in model accuracy, fault tolerance, robustness, and communication efficiency.