Source-linked AI summary
BrainTorrent: A Peer-to-Peer Environment for Decentralized Federated Learning
Abhijit Guha Roy, Shayan Siddiqui, Sebastian Pölsterl, Nassir Navab, Christian Wachinger
TL;DR
Medical centers need annotated data for deep learning, but pooling privacy-sensitive patient data is often infeasible. BrainTorrent addresses this by enabling server-less peer-to-peer federated learning, and achieves better performance than server-based federated learning in whole-brain segmentation, with gains up to 7% Dice points under unequal training-scan settings.
Problem
Medical centers often lack sufficiently large annotated datasets because annotation is expensive and time-consuming, while pooling patient data across centers is constrained by privacy.
Method
BrainTorrent is a server-less peer-to-peer federated learning environment in which medical centers communicate directly without a central server or sharing data.
Results
BrainTorrent achieves better performance than server-based federated learning across experimental settings, with gains up to 7% Dice points when clients have unequal numbers of training scans.
Takeaways & Limitations
Server-less BrainTorrent supports decentralized whole-brain segmentation training without relying on a central server body for orchestration.
Takeaways & Limitations
The study is a proof-of-concept evaluation of whole-brain segmentation with sensitive patient-specific data that cannot be shared across centers or with a central server.
Abstract
from arXiv · showhide
Access to sufficient annotated data is a common challenge in training deep neural networks on medical images. As annotating data is expensive and time-consuming, it is difficult for an individual medical center to reach large enough sample sizes to build their own, personalized models. As an alternative, data from all centers could be pooled to train a centralized model that everyone can use. However, such a strategy is often infeasible due to the privacy-sensitive nature of medical data. Recently, federated learning (FL) has been introduced to collaboratively learn a shared prediction model across centers without the need for sharing data. In FL, clients are locally training models on site-specific datasets for a few epochs and then sharing their model weights with a central server, which orchestrates the overall training process. Importantly, the sharing of models does not compromise patient privacy. A disadvantage of FL is the dependence on a central server, which requires all clients to agree on one trusted central body, and whose failure would disrupt the training process of all clients. In this paper, we introduce BrainTorrent, a new FL framework without a central server, particularly targeted towards medical applications. BrainTorrent presents a highly dynamic peer-to-peer environment, where all centers directly interact with each other without depending on a central body. We demonstrate the overall effectiveness of FL for the challenging task of whole brain segmentation and observe that the proposed server-less BrainTorrent approach does not only outperform the traditional server-based one but reaches a similar performance to a model trained on pooled data.
1 Introduction
Medical DNN training needs abundant annotated data, but annotation burdens and data-sharing restrictions leave individual centers with limited datasets. BrainTorrent addresses this decentralized setting with server-less peer-to-peer federated learning for medical centers.
- Motivation: Manual labeling of a single 3D brain MRI scan can take up to a week, making abundant annotations difficult to obtain.The paper identifies annotation cost and time as barriers to building sufficiently large in-house datasets.
- Motivation: Ethical and legal regulations restrict sharing medical data across centers or with a central server.Consequently, centers may each have too little data for effective DNN training while pooled training remains unavailable.
- Prior limitations: Sequential fine-tuning across centers can cause catastrophic forgetting as new data overwrites knowledge acquired from earlier training.The paper connects this effect to deteriorating generalizability.
- Federated learning: Federated learning coordinates decentralized training through a central server while avoiding direct sharing of client data.The paper notes that this strategy was introduced for decentralized DNN training and initially motivated by mobile-device collaboration.
- BrainTorrent: BrainTorrent is a server-less peer-to-peer federated learning environment in which medical-center clients communicate directly.Its design removes the trusted central body, supports dynamic client-initiated updates, and is intended for collaborative medical training.
- Scope and contribution: The paper applies federated learning to whole-brain segmentation with 20 classes and severe class imbalance, reporting BrainTorrent's advantage over server-based FL in exemplar heterogeneous settings.The authors describe this as the first application of FL to whole-brain segmentation.
2 Method
Traditional federated learning uses a central server to aggregate partially trained client models and redistribute the weighted result. BrainTorrent replaces that coordinator with direct peer-to-peer communication, version tracking, selective model exchange, and client-side aggregation.
- Federated Learning with Server: Traditional FL trains client models locally for a few iterations, sends their parameters to a central server, and redistributes the aggregated model.The server aggregates client weights by data-size-weighted averaging.
- Federated Learning with Server: The server model is redistributed across clients, while each client ultimately retains a personalized model fine-tuned to local data.A newly added client receives the server model as its starting point.
- BrainTorrent: BrainTorrent connects clients directly and maintains, alongside each model, a version vector recording its own version and versions used during merging.Version entries start at zero and the client's own version increases after each fine-tuning step.
- BrainTorrent: Each BrainTorrent round randomly selects one client to ping peers and obtain their latest model versions.This determines which peer updates are newer than the selected client's recorded versions.
- BrainTorrent: Only clients with newer updates send their weights and training-sample sizes to the initiating client for merging.The initiating client combines the received subset with its current model by weighted averaging.
- Algorithm: The algorithm initializes random client models and zero version vectors before repeatedly selecting clients, pinging peers, receiving eligible updates, and merging weights.These steps operationalize decentralized training without a coordinating server.
- Round definition: A BrainTorrent round updates one random client, unlike an FLS round that updates all clients once.Thus, R FLS rounds correspond to R × N BrainTorrent rounds in client-update count.
3 Experimental Settings
The experiments evaluate BrainTorrent on whole-brain MRI T1 segmentation under varying client counts, client-specific data, and non-uniform age-range distributions. Across these settings, BrainTorrent is compared with server-based federated learning and pooled-data models.
- Experimental setup: The study uses 30 annotated whole-brain MRI T1 scans, with 20 for training and 10 for testing, using QuickNAT for segmentation.The 20 training scans are distributed among clients, and the experiments vary the client count between 5, 7, 10, and 20.
- Experiment 1: Number of clients: With 5–20 clients, BrainTorrent outperforms FLS in both average client Dice score and aggregated-model Dice score.The comparison uses the pooled model as an upper-bound reference and evaluates the aggregated model supplied to a new client.
- Experiment 1: Client-wise performance: With 10 clients and 2 scans per client, BrainTorrent and FLS improve over client-only models by average gains of 26% and 24% Dice points, respectively.BrainTorrent also achieves an average 2% higher Dice score than FLS, with 7 of 10 client models performing better.
- Experiment 2: Non-uniform data: The second experiment distributes 20 training scans across 5 clients with non-overlapping age ranges and unequal numbers of scans.This setting represents client-specific characteristics and non-uniform data distribution in a clinical-use scenario.
- Experiment 2: Non-uniform data: Under uneven data distribution, BrainTorrent’s aggregated model matches the pooled model, while FLS is 3% Dice points lower and BrainTorrent leads FLS by 7% Dice points across clients.Clients with only 2 and 1 annotated scans perform comparatively poorly with FLS but very well with BrainTorrent.
4 Conclusion
The paper concludes that BrainTorrent is a server-less peer-to-peer federated learning environment for decentralized training. In whole-brain segmentation experiments, it outperforms server-based federated learning across settings, with performance similar to pooled-data training.
- Conclusion: BrainTorrent removes the central server from federated learning while supporting decentralized training of a fully convolutional segmentation network.The proof-of-concept study focuses on whole-brain segmentation.
- Conclusion: BrainTorrent outperforms server-based federated learning under different experimental settings, with the margin reaching 7% Dice points for unequal training-scan counts.Its performance is reported as similar to a model trained on pooled data across clients.
- Conclusion: Although demonstrated for image segmentation, the authors state that BrainTorrent can be used for training any machine learning model.