Source-linked AI summary
Multi-Institutional Deep Learning Modeling Without Sharing Patient Data: A Feasibility Study on Brain Tumor Segmentation
Micah J Sheller, G Anthony Reina, Brandon Edwards, Jason Martin, Spyridon Bakas
TL;DR
Medical-imaging collaboration needs more data for expert-labeled segmentation, yet centralized sharing faces legal, privacy, technical, and ownership barriers. The paper evaluates federated learning, which trains across institutions without sharing patient data, and finds performance close to centralized training while outperforming alternative collaborative approaches.
Problem
Medical imaging lacks sufficient data for expert-labeled segmentation, while centralized multi-institutional sharing faces legal, privacy, technical, and data-ownership challenges.
Method
The study applies federated learning to BraTS brain-tumor segmentation, aggregating locally trained models across institutions without sharing patient data, and compares IIL and CIIL.
Results
99% of centralized model performance is achieved by federated learning despite imbalanced datasets or simulations with 32 institutions and approximately 6 subjects per institution.
Takeaways & Limitations
Federated learning supports multi-institutional clinical model training while retaining patient data locally and avoiding the need for centralized data sharing.
Takeaways & Limitations
Differentially private segmentation training is not evaluated because update noise can slow training and leave the model below the desired utility.
Abstract
from arXiv · showhide
Deep learning models for semantic segmentation of images require large amounts of data. In the medical imaging domain, acquiring sufficient data is a significant challenge. Labeling medical image data requires expert knowledge. Collaboration between institutions could address this challenge, but sharing medical data to a centralized location faces various legal, privacy, technical, and data-ownership challenges, especially among international institutions. In this study, we introduce the first use of federated learning for multi-institutional collaboration, enabling deep learning modeling without sharing patient data. Our quantitative results demonstrate that the performance of federated semantic segmentation models (Dice=0.852) on multimodal brain scans is similar to that of models trained by sharing data (Dice=0.862). We compare federated learning with two alternative collaborative learning methods and find that they fail to match the performance of federated learning.
1 Introduction
Glioma segmentation can support quantitative analysis of tumor heterogeneity and links to disease characteristics, but centralized multi-institutional datasets face major practical and governance barriers. This study introduces federated learning for medical imaging, training across institutions without sharing patient data and comparing it with alternative collaborative methods.
- Accurate segmentation of glioma subregions supports quantitative imaging features linked to disease properties, treatment response, prognosis, and molecular characteristics.
- Centralized medical-imaging collaboration is constrained by limited data availability and legal, privacy, technical, and data-ownership challenges.
- Federated learning trains a shared segmentation model by iteratively aggregating locally trained models while institutions retain patient data.
- The study compares federated learning with institutional incremental learning and cyclic institutional incremental learning.
- IIL performs poorly relative to FL and CIIL, while CIIL is less stable and harder to validate than FL.
2 Materials and Methods
The study uses federated learning to train a U-Net segmentation model across distributed medical-imaging institutions, evaluating convergence settings, alternative collaborative methods, privacy considerations, and the BraTS dataset. The task focuses on whole-tumor segmentation from FLAIR MRI in high-grade glioma patients.
- Federated learning: Federated learning keeps patient data locally, sends model updates to a central server, and aggregates institutional updates into shared parameters.The server returns updated parameters for further local training or application.
- Hyper-Parameters and Convergence: Federated learning includes epochs per round, participant count per round, and model-update compression or pruning as additional hyper-parameters.This study varies epochs per round and federation size, but does not assess participant count or compression/pruning.
- Privacy considerations: Differential privacy can obscure information in model updates, but noise may slow training and leave utility below the desired level; private segmentation training is left for future work.
- Collaborative learning alternatives: IIL trains a shared model sequentially across institutions, whereas CIIL repeats this process with fixed local epoch counts to reduce forgetting.FL instead trains institutions in parallel and aggregates their updates after each round.
- U-Net model: The U-Net takes a single-channel image and outputs an equivalently sized binary mask, using contracting and expanding paths with skip connections.
- BraTS Dataset: The BraTS 2018 training dataset contains multi-institutional, multimodal MRI brain scans from glioma patients with manually annotated radiographic abnormalities.
- BraTS Dataset: The evaluation targets whole-tumor volume in high-grade glioma patients using only the FLAIR modality, rather than developing a new segmentation method.
3 Experimental Results
The experiments evaluate centralized data-sharing and three collaborative learning methods across real and simulated institutional distributions. Federated learning closely matches centralized performance, while IIL and CIIL show instability or lower performance, with FL remaining effective as the number of institutions increases.
- 3.2 Baseline U-Net Results: 0.862 validation DC was reached within 3 epochs for the baseline U-Net trained on a 15% holdout configuration.The model reached state-of-the-art accuracy within 3 epochs and peaked at validation DC 0.862.
- 3.3 BraTS Distribution Results: CIIL instability required evaluation after each cycle, adding model-transfer and result-aggregation overhead.These requirements made CIIL arguably more complex than FL and increased communication cost.
- 3.3 BraTS Distribution Results: 98.7% of centralized validation DC was achieved by FL on the actual BraTS distribution despite substantial institutional imbalance.The best CIIL results also reached 98.7%, but CIIL was less stable across cycles.
- 3.3 BraTS Distribution Results: IIL averaged only 93% of validation DC and showed instability similar to CIIL.IIL trained institutions sequentially, passing the best-performing model to the next institution after local validation stopped improving.
- 3.4 Results for Random Simulated Distributions: 10x larger standard deviations than FL were observed for CIIL and IIL with 16 and 32 simulated institutions.FL achieved 99+% of data-sharing results in all simulated distributions, including federations averaging fewer than 6 subjects per institution at 32 institutions.
- 3.4 Results for Random Simulated Distributions: Similar model quality was reached across FL federation sizes, but convergence rates differed as institutions became more numerous.Smaller datasets produced smaller model updates and less similar per-institution datasets; smaller model deltas caused a convergence slowdown.
4 Practical considerations
Practical deployment depends on accounting for acquisition variability and reproducible labeling. The study also considers how collaborator turnover may affect what the model retains and learns.
- Variation in equipment configurations and acquisition protocols must be considered because clinical acquisition environments are uncontrolled.More controlled environments, such as clinical trials, are described as more suitable for large-scale analytical studies.
- Reproducible labeling requires documented semantic descriptors of distinct anatomical regions across institutions.The labeling protocol links training data to reproducible expert clinical knowledge while avoiding operator bias.
- 4.3 Addition/Removal of Collaborators: Adding or removing institutions is expected to produce a qualitatively similar model after a transition period, but missing data would eventually be forgotten.New data patterns would be learned subject to the limitations of the selected collaborative configuration and new data distribution.
5 Conclusions
The experiments show that federated learning can train segmentation models across clinical institutions without sharing patient data. Its performance remains near data-sharing models, while alternative approaches face communication or scalability disadvantages.
- Federated learning achieved 99% of data-sharing model performance with imbalanced data and with 32 institutions averaging six subjects each.The conclusion reports this result across both the actual BraTS institutional distribution and the 32-institution simulation.
- CIIL requires frequent full validation to select a good model, adding synchronization, aggregation, and communication costs above federated learning.These validation steps make CIIL more complex than FL in the reported comparison.
- IIL and CIIL do not scale well when many institutions have small amounts of data.
- Clinical adoption of federated learning is expected to support multi-institutional computer-aided analytics and assistive diagnostics without sharing patient data.The paper connects this prospect to addressing potential privacy or data-ownership concerns.