Source-linked AI summary
Multi-site fMRI Analysis Using Privacy-preserving Federated Learning and Domain Adaptation: ABIDE Results
Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence Staib, Pamela Ventola, James S. Duncan
TL;DR
Multi-site fMRI analysis needs large datasets, but medical institutions face data-sharing constraints and heterogeneous scans. This paper develops privacy-preserving federated learning with domain adaptation and finds advantages over alternative methods, while identifying possible ASD biomarkers.
Problem
Large-scale, generalizable fMRI models are difficult to develop because single sites lack sufficient data and scans differ across institutions.
Method
The paper trains multi-site fMRI models through federated learning with randomized shared weights and domain adaptation for heterogeneous site distributions.
Results
Federated learning showed advantages over alternative methods, domain adaptation potentially improved performance, and the model revealed possible brain biomarkers for identifying ASD.
Takeaways & Limitations
The framework supports using multi-site data without data sharing and may help identify ASD biomarkers, with implications for diseases involving fewer patients.
Takeaways & Limitations
The study could not explicitly provide the privacy bound ϵ because the deep learning classifier’s sensitivity was difficult to estimate.
Abstract
from arXiv · showhide
Deep learning models have shown their advantage in many different tasks, including neuroimage analysis. However, to effectively train a high-quality deep learning model, the aggregation of a significant amount of patient information is required. The time and cost for acquisition and annotation in assembling, for example, large fMRI datasets make it difficult to acquire large numbers at a single site. However, due to the need to protect the privacy of patient data, it is hard to assemble a central database from multiple institutions. Federated learning allows for population-level models to be trained without centralizing entities' data by transmitting the global model to local entities, training the model locally, and then averaging the gradients or weights in the global model. However, some studies suggest that private information can be recovered from the model gradients or weights. In this work, we address the problem of multi-site fMRI classification with a privacy-preserving strategy. To solve the problem, we propose a federated learning approach, where a decentralized iterative optimization algorithm is implemented and shared local model weights are altered by a randomization mechanism. Considering the systemic differences of fMRI distributions from different sites, we further propose two domain adaptation methods in this federated learning formulation. We investigate various practical aspects of federated model optimization and compare federated learning with alternative training strategies. Overall, our results demonstrate that it is promising to utilize multi-site data without data sharing to boost neuroimage analysis performance and find reliable disease-related biomarkers. Our proposed pipeline can be generalized to other privacy-sensitive medical data analysis problems.
1. Introduction
The introduction motivates multi-site fMRI analysis by balancing the benefits of pooling scarce medical data against privacy risks and site-specific distribution differences. It presents privacy-preserving federated learning with domain adaptation as the paper’s proposed response.
- Motivation: Pooling large-scale medical data could improve the generalizability, accuracy, and reproducibility of deep-learning models, but comparable healthcare datasets are difficult to assemble.The introduction notes that image-recognition models may require training sets of up to ten thousand images, while healthcare data at scale are often unavailable.
- Motivation: Medical data sharing raises patient, anonymity, and legal concerns, including possible use by employers or insurers even after deidentification.Patients may worry that data stored and accessed by multiple users could affect future employment or health-insurance decisions.
- Federated learning: Federated learning addresses these concerns by keeping data at each institution while enabling decentralized computation and multi-institution collaboration.The approach avoids transferring data to a centralized warehouse for model building.
- Domain adaptation: Heterogeneous scanners and participant instructions create site-specific fMRI distribution differences, motivating domain adaptation for multi-site analysis.The cited sites used different 3T scanners and gave participants different instructions.
- Contributions: The paper contributes a privacy-preserving multi-site fMRI federated-learning pipeline, examines communication and privacy mechanisms, and investigates domain adaptation in federated medical-image analysis.The introduction identifies domain shift from heterogeneous institutional data distributions as a central challenge.
2. Related Work
Related work frames federated learning as privacy-preserving collaboration without sharing patient data, while domain adaptation transfers knowledge from source to target domains. The paper focuses on parameter-sharing federated learning for medical imaging and builds on established domain-adaptation research.
- Federated Learning: Federated learning uses either private local training with model-parameter transfer or encryption techniques for secure communication between parties.The paper focuses on the first approach, studied in prior work by Dean et al., Shokri and Shmatikov, and McMahan et al.
- Medical Imaging Collaboration: Medical imaging requires substantial data and expert labeling, motivating institutional collaboration while legal and technical issues constrain centralized sharing.These challenges are especially pronounced when sharing medical data among international institutions.
- Domain Adaptation: Domain adaptation transfers knowledge learned from a source domain to a target domain by refining a source-trained model on different target-domain data.The passage identifies unsupervised domain adaptation as an extensively studied research area.
3. Methods
The methods formulate multi-site fMRI classification as privacy-preserving federated learning, keeping institutional data local while collaboratively training a shared model. The framework uses decentralized local updates and periodic global communication, with Gaussian or Laplace noise added to shared weights for privacy protection.
- Federated learning formulation: The multi-site setting is horizontal federated learning because sites share feature and label spaces while having limited sample overlap.The formulation specifies Xi = Xj, Yi = Yj, and Ii ∩ Ij as empty for distinct sites.
- Federated learning formulation: The framework models N institutions as data owners that collaboratively train a shared deep learning architecture without exposing local fMRI data to other parties.A central server performs computation but does not store data, and each site uses the same architecture for the same task.
- Federated optimization: Training alternates between local model updates and communication with a global server, with the objective at each site defined by cross-entropy loss.The global model is periodically deployed back to local models according to the communication pace τ, and Algorithm 1 runs for K optimization iterations.
- Privacy protection: Differential privacy controls the privacy preference through ϵ, which bounds what an attacker can learn about an individual from model outputs.The additive term δ denotes the probability that ϵ-differential privacy is broken.
- Privacy protection: The privacy-preserving mechanism adds noise to shared model weights using either the Gaussian mechanism or the Laplace mechanism.Local parameters are randomized at initialization, and the noise generator M(·) is used for privacy preservation during federated training.
Gaussian Mechanism
The Gaussian mechanism adds zero-mean Gaussian noise calibrated to a function’s global sensitivity, linking the noise scale σ to (ϵ, δ)-differential privacy. Under the stated condition, this mechanism provides privacy for ϵ < 1.
- Gaussian Mechanism: δ ≥ 4 5exp(−(σϵ)2/2) ensures (ϵ, δ)-differential privacy when ϵ < 1.This condition applies to the Gaussian mechanism under global sensitivity sh.
- Gaussian Mechanism: The formulation links the Gaussian noise parameter σ to the privacy parameters ϵ and δ.The privacy guarantee is expressed through the stated relationship among σ, ϵ, and δ.
- Gaussian Mechanism: The Gaussian mechanism adds Gaussian noise with mean 0 and standard deviation shσ to a function h(D) with global sensitivity sh.The noise is described as N(0, s2 hσ2).
Laplace Mechanism
The privacy mechanism adds Laplace noise calibrated to the model’s global sensitivity and privacy parameter ε, with sensitivity assumed to be 1 for tractability. The section also describes domain adaptation methods that align site-specific distributions while preserving locally stored data.
- Laplace Mechanism: The Laplace mechanism adds Lap(sh/ε) noise to h(D) and preserves (ε, 0)-difference privacy.The Laplace distribution has variance σ2 = 2b2, linking noise scale b to privacy parameter ε.
- Laplace Mechanism: Because h is a deep learning model with intractable sensitivity, the method assumes sh = 1 to control noise for a required privacy level.Noise parameters are linked to privacy parameters through the mechanism.
- Domain Adaptation: Domain adaptation is motivated by site-specific data distributions, with the hypothesis that it can improve federated accuracy even when privacy noise is added.The expected benefit is especially relevant for sites whose data distributions differ.
- Mixture of Experts: The federated Mixture of Experts combines a collaboratively learned general model with private domain experts using learned gating weights.Each site retains domain-adapted expert models while jointly building a general task model.
- Federated Adversarial Alignment: Federated adversarial alignment uses local generators, classifiers, and discriminators to minimize discrepancy between source and target domains without jointly accessing both datasets.The alignment module is trained end-to-end within the federated model, with implementation specified in Algorithm 2.
4. Experiments and Results
Experiments on ABIDE evaluated federated ASD-versus-healthy-control classification, privacy noise, alternative training strategies, and domain adaptation. Results showed privacy–performance trade-offs, site-dependent adaptation benefits, and robust federated biomarkers across four sites.
- Experimental setup: The study classified autism spectrum disorder versus healthy control using ABIDE I resting-state fMRI functional-connectivity features and a 6105-16-2 MLP with subject-wise 5-fold cross-validation.Functional connectivity was computed from ROI mean time sequences using Pearson correlations, Fisher-transformed before classification.
- Communication pace: Communication pace τ from 5 to 30 produced no significant accuracy differences when shared-weight noise was omitted.The experiment examined communication cost while selecting τ.
- Training-strategy comparison: Using τ = 20 and α = 0.01, federated learning was compared with Single, Cross, Mix, and Ensemble strategies for multi-site classification.Mix used all sites’ data but did not preserve privacy, whereas the federated approach avoided data sharing.
- Domain adaptation: Fed-MoE showed potential to improve UCLA accuracy over Fed (p = 0.094), while domain adaptation did not improve USM performance, likely because of its data distribution.The heterogeneous site distributions motivated testing federated domain-adaptation methods.
- Biomarker analysis: Fed detected robust classification biomarkers across 4 sites, whereas Single identified different biomarkers across sites.The study plotted the top 10 ROI importance scores for healthy-control and autism-spectrum-disorder classifications.
5. Conclusion
The paper presents a privacy-preserving federated learning framework for multi-site fMRI analysis that addresses domain shift and demonstrates the potential of collaborative modeling without data sharing. The framework may also identify ASD-related brain biomarkers and generalize to other privacy-sensitive medical applications.
- Conclusion: The framework combines federated learning with randomized privacy-preserving mechanisms for multi-site fMRI classification.It uses brain functional connectivity to classify ASD and HC while protecting local data.
- Conclusion: Two domain adaptation strategies, MoE and adversarial domain alignment, were proposed to address domain shift and potentially improve federated model performance.The paper also discusses the conditions under which domain adaptation provides benefits.
- Conclusion: Federated learning can utilize multi-site data without data sharing and outperform alternative training methods.The results support collaborative model training while preserving institutional data isolation.
- Conclusion: The federated model revealed possible brain biomarkers for identifying ASD.This finding suggests the approach can support disease-related neuroimaging analysis.
- Conclusion: The approach establishes a unified model across medical institutions while protecting local data and may extend to other disease areas, particularly rare diseases.It is intended to accelerate privacy-sensitive deep learning applications in medical imaging.
Appendix · Architecture of the models
The appendix defines notation for the model architectures and directs readers to tables describing the Single, Cross, Ensemble, Fed, Mix, Fed-MoE, and Fed-Align strategies for ABIDE rs-fMRI classification.
- Architecture of the models: The architecture description specifies input and output dimensions for fully connected layers and zeroing probabilities for Dropout layers.Batch normalization, ReLU, and Softmax layers are denoted as BN, ReLU, and Softmax, respectively.
- Architecture of the models: Models for the Single, Cross, and Ensemble strategies are presented in Table 5.Table 5 is labeled as covering the ABIDE rs-fMRI classification task.
- Architecture of the models: The text states that models for Cross and Ensemble are shown in Table 6.This statement differs from Table 6’s caption, which identifies the Fed and Mix strategies.
- Architecture of the models: Table 6 is captioned as presenting model architectures for the Fed and Mix strategies in ABIDE rs-fMRI classification.The caption identifies the task and the two strategies covered by the table.
- Architecture of the models: The Fed-MoE strategy model architecture is presented in Table 7.Table 7 is identified as covering the ABIDE rs-fMRI classification task.
- Architecture of the models: The Fed-Align strategy model architecture is presented in Table 8.The passage provides the table reference but no additional architectural specifications.
Names of the biomarkers
The section lists the top 10 important brain regions identified as HC or ASD biomarkers by Fed and Single analyses across NYU, UM, USM, and UCLA. Fed-derived HC and ASD lists prominently include temporal, frontal, occipital, subcortical, and cingulate regions, while Single-derived lists are also site-specific.
- HC biomarkers detected by Fed: Fed identifies the top 10 important ROIs as HC biomarkers across NYU, UM, USM, and UCLA.The listed regions include Right Heschl’s Gyrus, temporal gyri, frontal gyri, the Left Intracalcarine Cortex, cingulate regions, and the Left Planum Polare.
- HC biomarkers detected by Single: Single identifies site-specific HC biomarker lists, including temporal, occipital, supramarginal, paracingulate, precentral, frontal-orbital, and parahippocampal regions.NYU, UM, USM, and UCLA lists contain overlapping but nonidentical regions.
- ASD biomarkers detected by Fed: Fed identifies ASD biomarkers including the Left Accumbens, thalamic, pallidal, frontal, precentral, occipital, opercular, and temporal-fusiform regions across sites.The site lists share several regions but vary in ordering and composition.
- ASD biomarkers detected by Single: Single ASD biomarker lists contain site-specific combinations of occipital-fusiform, angular, putamen, thalamic, cingulate, opercular, pallidal, and parahippocampal regions.The listed regions differ among NYU, UM, USM, and UCLA.
Additional experiments using alternative atlas
The authors replaced the structural Harvard-Oxford atlas with the functional Craddock 200 atlas and replicated the main experiment using the resulting 19,900-dimensional inputs. CC200 produced lower classification accuracies than HO overall.
- Atlas replacement: The structural Harvard-Oxford atlas was replaced with the functional Craddock 200 atlas, which parcellates the brain into 200 regions of interest.The preprocessing pipeline from Section 4.1.2 was followed.
- Experimental setup: Replicating the main experiment with CC200 changed each MLP’s input dimension from 6105 to 19900 (= 200 × 100 −100).The comparison results were reported in Appendix Table 9.
- Results: Overall, classification accuracies using CC200 were lower than those using HO.Appendix Table 9 reports mean classification accuracies and corresponding standard deviations for different training strategies with CC200.
Additional experiments on ASD sex classification
The paper extends its federated-learning evaluation to ASD sex classification and analyzes the resulting brain biomarkers. Federated strategies improved over random guessing, while the analysis highlighted Frontal Gyrus and Angular Gyrus effects.
- Additional experiments on ASD sex classification: The study replicated its main experiment for ASD sex classification, excluding USM because it contained only male ASD subjects.The random-guess accuracy was used as the baseline strategy.
- Additional experiments on ASD sex classification: Federated-related strategies showed obvious improvement over the random-guess baseline in ASD sex classification.
- Biomarker interpretation: Gradient-based feature-importance analysis aggregated functional-connectivity scores into normalized importance scores for 111 ROIs.The method formed a symmetric gradient matrix, summed it over columns, normalized the resulting vector to [0, 1], and averaged scores by site.
- Biomarker interpretation: For male ASD classification, the methods highlighted Frontal Gyrus and Angular Gyrus ROIs, consistent with previously reported sex effects.The top 10 important ROIs and their normalized scores were plotted for NYU, UM, and UCLA sites.