dcSBM: A federated constrained source-based morphometry approach for multivariate brain structure mapping

Debbrata K. Saha, Rogers F. Silva, Bradley T. Baker, Rekha Saha, Vince D. CalhounbioRxiv 2022 · DOI

TL;DR

This paper introduces decentralized constrained source-based morphometry (dcSBM), a novel method for analyzing brain structure across multiple sites without pooling sensitive patient data. By performing constrained Independent Component Analysis (ICA) locally and aggregating results, dcSBM achieves comparable accuracy to traditional centralized methods, enabling broader collaboration in neuroimaging research.

What Problem Does This Paper Address?

Analyzing brain morphometry using structural MRI (sMRI) is crucial for understanding neurological differences between groups (e.g., patients vs. controls). Source-based morphometry (SBM) and its constrained variant (constrained SBM) are powerful data-driven techniques for identifying these patterns. However, a major limitation of constrained SBM is its requirement for local data accessibility. This poses a significant challenge in multi-site studies due to privacy concerns (e.g., rare disease information), strict Institutional Review Board (IRB) policies, and the logistical/financial burden of data sharing. Consequently, valuable decentralized data often goes underutilized. This paper addresses this gap by proposing a decentralized approach that allows constrained SBM to operate effectively without centralizing raw data.

Methods

The core innovation is decentralized constrained source-based morphometry (dcSBM). This approach extends constrained SBM to a distributed environment. Here's how it works:

  • Local Constrained ICA: Each participating site performs constrained Independent Component Analysis (ICA) on its private local data. This process is guided by a shared reference map (R) that encodes prior information about desired sources. The constrained ICA algorithm, implemented using a fast fixed-point algorithm from the GIFT toolbox, decomposes the local data into mixing matrices (Aℓ, representing subject contributions to sources) and source matrices (Sℓ, representing spatial maps of sources).
  • Decentralized Aggregation: After local processing, the mixing matrices (Aℓ) from each site are aggregated by a master node. The source matrices (Sℓ) are also aggregated for quality control purposes.
  • Statistical Analysis: The aggregated results are then used for statistical analysis, allowing for comparisons between centralized and decentralized estimates.

This approach leverages the multivariate and exploratory power of constrained SBM while preserving data privacy by ensuring that raw data never leaves its local site.

Datasets

The authors evaluated their dcSBM approach using three datasets:

  • UK Biobank: A large sample of 3000 unaffected subjects' T1-weighted sMRI data was used to investigate the reliability of the dcSBM algorithm.
  • Function Biomedical Informatics Research Network (fBIRN): This multi-site dataset, collected from seven different sites, was used to validate the model by comparing centralized and decentralized results. It included 165 healthy controls and 152 patients with schizophrenia.
  • Centers of Biomedical Research Excellence (COBRE): This dataset, collected from a single site, was also used for validation. It included 82 healthy controls and 66 patients with schizophrenia.

Key Insights and Ideas

  • Privacy-Preserving Neuroimaging Analysis: The most significant insight is the successful development of a method that enables sophisticated neuroimaging analysis (constrained SBM) without compromising data privacy or requiring data pooling. This opens doors for collaborative research across institutions with strict data sharing policies.
  • Decentralized ICA is Viable: The paper demonstrates that performing ICA in a decentralized manner, specifically constrained ICA, can yield results highly comparable to traditional centralized methods. This suggests that distributed computation can be effectively applied to complex neuroimaging analysis pipelines.
  • Reliability Across Sites: The study highlights the reliability of dcSBM, showing consistent results across different sites and datasets, even when simulating decentralized scenarios. This is crucial for building trust in decentralized analytical frameworks.
  • Clinical Relevance: By applying dcSBM to patient datasets (fBIRN and COBRE), the authors show that the method can identify clinically relevant patterns, such as differences associated with schizophrenia, demonstrating its practical utility beyond theoretical validation.

Key Results

  • High Correlation: The study found a high correlation between centralized and decentralized loading parameters and source matrices across all experiments. For instance, Figure 2(B) shows a strong diagonal correlation between centralized and decentralized loading parameters, indicating that the identified components are highly similar.
  • Similar Effect Sizes and P-values: Regression analysis revealed that the effect sizes and p-values for variables like age, gender, and diagnosis were very similar between centralized and decentralized analyses (Figures 4, 6, and Table 1 & 2). This consistency reinforces the validity of the dcSBM approach.
  • Model Reliability: Experiments using the UK Biobank dataset (Figure 7 & 8) demonstrated high reliability, with performance metrics like correlation and Mean Square Error (MSE) remaining consistent across multiple shuffled runs and increasing numbers of sites. For example, Figure 7(A) shows high similarity metrics for loading correlations across 10 shuffled runs.
  • Identification of Clinically Relevant Sources: The analysis identified several independent sources with large effect sizes, particularly for diagnosis. For example, components 4 and 28 were highlighted for their significant association with diagnosis, covering brain regions known to be involved in psychiatric disorders.

Limitations and Future Work

  • Scaling Ambiguity in ICA: The authors acknowledge that estimated loading parameters from dcSBM can have larger values than their centralized counterparts, a known issue with ICA scaling ambiguity. While they modeled site effects to mitigate this, it remains a potential concern for site effect inflation.
  • Aggregation for Quality Control: While the source matrices were aggregated for quality control, the authors note that retaining them locally at each site would further minimize privacy risks.
  • Future Work:
    • Decentralized Regression: The authors plan to deploy a decentralized regression approach that avoids aggregating results altogether, further enhancing privacy.
  • Integration with COINSTAC: They aim to integrate dcSBM within the COINSTAC platform to make it openly available to the research community.
  • Further Validation: Continued validation with more diverse datasets and clinical populations is implied.

Summary updated

Abstract

Abstract The examination of multivariate brain morphometry patterns has gained attention in recent years, especially for their powerful exploratory capabilities in the study of differences between patients and controls. Among many existing methods and tools for analysis of brain anatomy based on structural magnetic resonance imaging (sMRI) data, data-driven source based morphometry (SBM) focuses on the exploratory detection of such patterns. Constrained source-based morphometry (constrained SBM) is a widely used semi-blind extension of SBM that enables extracting maximally independent reference-alike sources using the constrained independent component analysis (ICA) approach. In order to operate, constrained SBM needs the data to be locally accessible. However, there exist many reasons (e.g., the concerns of revealing identifiable rare disease information, or violating strict IRB policies) that may preclude access to data from different sites. In this scenario, constrained SBM fails to leverage the benefits of decentralized data. To mitigate this problem, we present a novel approach: decentralized constrained source-based morphometry (dcSBM). In dcSBM, the original data never leaves the local site. Each site operates constrained ICA on their private local data while using a common distributed computation platform. Then, an aggregator/master node aggregates the results estimated from each local site and applies statistical analysis to find out the significant sources. In our approach, we first use UK Biobank sMRI data to investigate the reliability of our dcSBM algorithm. Finally, we utilize two additional multi-site patient datasets to validate our model by comparing the resulting group difference estimates from both centralized and decentralized constrained SBM.
BibTeX
@misc{saha2022dcsbm,
  author = {Debbrata K. Saha and Rogers F. Silva and Bradley T. Baker and Rekha Saha and Vince D. Calhoun},
  title = {{dcSBM: A federated constrained source-based morphometry approach for multivariate brain structure mapping}},
  year = {2022},
  doi = {10.1101/2022.12.29.522266},
  howpublished = {bioRxiv},
}
Figure