Source-linked AI summary
Bayesian Uncertainty Quantification for fMRI Functional Connectivity via Simulation-Based Inference
Simon Carter, Zeming Kuang, Lilianne R. Mujica-Parodi, Helmut H. Strey
TL;DR
Traditional connectivity methods do not separate scanner noise from between-subject neural variability, limiting uncertainty estimates for scan planning and individual measurements. The paper models BOLD dynamics and measurement noise with a Bayesian framework and applies simulation-based inference. It reports a population-level precision advantage for 7T over 3T, while noting limits to attributing cross-dataset differences specifically to field strength.
Problem
Conventional connectivity measures cannot separate scanner measurement noise from true between-subject neural variability, and researchers lack principled uncertainty bounds for assessing scan duration and individual differences.
Source: §1 · p. 3Method
Results
7T maintains a population-level advantage over 3T, with the paper reporting that the advantage grows stronger at the population level.
Source: §7.1 · p. 16, §7.6 · p. 19Takeaways & Limitations
The framework provides a tool for optimizing scan duration, spatial resolution, and group size for a specific scanner and population.
Source: §7.1 · p. 16Field-strength effects cannot be disentangled from population, acquisition-protocol, and preprocessing differences without a controlled comparison; projected population curves also rely on 1/√n contraction and regularity assumptions.
Source: §7.6 · p. 20
Abstract
from arXiv · showhide
Optimizing fMRI scan duration and spatial resolution is critical for experimental design, yet traditional correlation-based approaches cannot quantify uncertainty or disentangle scanner measurement noise from true neural variability across subjects. Without principled uncertainty bounds, researchers cannot know whether a protocol is long enough to reliably estimate connectivity, or whether between-subject differences reflect biological variation or noise. We present a Bayesian framework modeling BOLD dynamics as coupled Ornstein-Uhlenbeck processes, using Sequential Neural Posterior Estimation to obtain connectivity posteriors while accounting for frequency-independent measurement noise across the BOLD spectrum. Applied to N = 28 healthy controls (55 scans) at 7T using a functional network atlas (65 DMN regions), the framework quantifies uncertainty across its sources: scanner noise, subject variability, and acquisition length. Spatial analysis identifies a mean of 46 voxels per ROI, roughly half of typical region sizes, as sufficient to achieve 90% of asymptotic precision. At the single-subject level, 7T reaches its within-session precision plateau in approximately 7 minutes versus 10 minutes for 3T, a 40% reduction in required scan time, providing the first direct, model-based quantification of the scan-time advantage conferred by higher field strength. At the population level, 3T requires roughly 37 times more per-subject scan time than 7T for the pooled curves to converge, confirming a consistent advantage of higher field strength at every timescale. Together these findings provide concrete, scanner-specific guidance for protocol optimization, with direct implications for reducing acquisition costs and improving the reliability of connectivity-based clinical biomarkers. We provide code enabling researchers to derive these bounds from their own data.
1 Introduction
Empirical reliability benchmarks for connectivity depend on repeated scans or split-half designs, while conventional measures cannot separate scanner noise from biological variability. The paper proposes a generative Bayesian framework to estimate uncertainty from one acquisition, addressing the need for principled protocol and single-subject clinical decisions.
1 Introduction
- Empirical reliability benchmarks often require repeated sessions or split-half designs, and collecting the necessary test-retest data is costly.Existing studies report edgewise stabilization after approximately 5–15 minutes, with plateaus depending on acquisition and network factors.Source: §1 · p. 2
- Conventional connectivity measures cannot distinguish genuine between-subject neural variability from scanner measurement noise.This makes observed differences between subjects ambiguous.Source: §1 · p. 3
- Without principled uncertainty bounds, researchers cannot tell whether a protocol is long enough or whether apparent subject differences reflect biology or noise.This uncertainty constrains single-subject clinical biomarkers, where group averaging is unavailable.Source: §1 · p. 3
- The proposed Bayesian framework models BOLD dynamics with coupled Ornstein-Uhlenbeck processes and frequency-independent measurement noise to quantify uncertainty from one acquisition.It evaluates uncertainty from scanner noise, acquisition length, and between-subject variability.Source: §1 · p. 3
2 Coupled Ornstein-Uhlenbeck Model
The model represents regional BOLD dynamics as coupled Ornstein-Uhlenbeck processes and adds measurement noise to observed signals. A separate decoder maps model parameters to the underlying noiseless correlation, using summary statistics that capture regional dynamics and coupling.
2 Coupled Ornstein-Uhlenbeck Model
- Coupled Ornstein-Uhlenbeck processes model the underlying regional dynamics, with a scaling factor allowing observed signals to exhibit anticorrelation despite positive coupling.The model assigns positive coupling strength a synaptic interpretation and uses the scaling factor to determine whether the connection is excitatory or inhibitory.Source: §2 · pp. 3–4
- Independent Gaussian measurement noise is added to the latent regional signals, and the parameterization contains eight model parameters.These parameters include mean reversion, coupling, volatility, measurement noise, and signal scaling.Source: §2 · p. 4
- A separate decoder approximates noiseless zero-lag cross-correlation from the full parameter posterior because correlation depends nonlinearly on all eight parameters.The noiseless quantity excludes measurement-noise contributions.Source: §2 · p. 4
- The decoder is a fully connected MLP with architecture 8 →32 →16 →1, ReLU hidden-layer activations, and a linear output.It was trained with Adam and mean squared error for 50,000 epochs, then exported as TorchScript.Source: §2 · p. 4
- Seven summary statistics reduce the observed data while capturing individual process dynamics through autocorrelations and coupling through zero-lag cross-correlation.The autocorrelation features include lags of 1, 5, and 10 time steps.Source: §2 · pp. 4–5
- Simulated and observed signals receive the same fifth-order Butterworth bandpass filter from 0.01–0.1 Hz.The preprocessing is described as consistent with standard fMRI pipelines.Source: §2 · p. 5
3 SBI Methodology
The paper uses simulation-based inference to estimate connectivity uncertainty while explicitly modeling signal and measurement noise. Amortized Neural Posterior Estimation avoids repeated likelihood evaluation and supports rapid inference across many region pairs.
3 SBI Methodology
- Simulation-based inference models signal generation and noise to separate signal from noise and estimate individual-subject uncertainty beyond bandpass filtering and temporal averaging.The paper notes that residual noise remains after filtering and that trial averaging is problematic for single-subject clinical metrics.Source: §3 · p. 5
- Amortized Neural Posterior Estimation learns a mapping from observed-data summary statistics to model-parameter posteriors using simulations.A trained estimator performs inference with one forward pass, avoiding likelihood evaluation and per-query simulation.Source: §3 · p. 5
- Training used 500,000 prior simulations, with the posterior conditioned on seven summary statistics and a separately trained decoder predicting noiseless correlation.The trained posterior was serialized for conditional sampling without retraining.Source: §3 · pp. 5–6
4 Data and Preprocessing
The study analyzes a 7T resting-state dataset and uses HCP-YA data for a 3T comparison, applying a common atlas and filtering approach. The two subsections describe the datasets and their acquisition or preprocessing scope.
4.1 Bipolar Dataset (7T)
- The Bipolar dataset comprises 55 scans from 28 healthy controls acquired at 7T and parcellated into 65 default-mode-network regions.The scans used 1.75 mm isotropic voxels, TR 802 ms, and 730 measurements, approximately 10 minutes; scans with mean framewise displacement above 0.5 mm were excluded.Source: §4.1 · p. 6
4.2 HCP Young Adult Dataset (3T)
- HCP-YA resting-state data provide the 3T comparison, using the same Seitzman atlas and 0.01–0.1 Hz bandpass filtering as the 7T data.The passage identifies the data as standard HCP minimally preprocessed resting-state fMRI.Source: §4.2 · p. 6
5 Validation on Synthetic Data
Synthetic coupled Ornstein-Uhlenbeck simulations show that the model’s main advantage is uncertainty quantification, with modest point-estimation gains and conservative posterior intervals. Calibration was assessed by comparing empirical coverage with nominal credibility levels.
5 Validation on Synthetic Data
- The validation compared the SBI model with Pearson correlation using 1,000 simulated band-pass filtered coupled OU processes.The OU processes were selected as models of resting-state neuroimaging signals and capture stationary autocorrelation and coupling-induced correlations.Source: §5 · p. 7
- The SBI model was trained on 100,000 simulations, using uniform priors except for an exponential coupling prior chosen to yield approximately uniform signal correlations.Source: §5 · p. 8
- 1.5% overall and 3.42% for |ρ| ≥0.2 point-estimation improvements accompany the model’s primary advantage: full posterior distributions for uncertainty quantification.The posterior examples show distributions centered near the true correlations for both high- and low-correlation cases.Source: Figure 1 · p. 7, §5 · p. 8
- 0.1125 integrated calibration error accompanies empirical coverage consistently exceeding nominal credibility, indicating conservative uncertainty estimates.The calibration plot compares empirical coverage with credibility levels; its curve lies above the diagonal reference.Source: Figure 1 · p. 7, §5 · p. 8
6 Validation on Resting-State Data
Resting-state validation examines how uncertainty estimates can guide spatial resolution and scan duration for DMN connectivity, at both single-subject and population levels.
6 Validation on Resting-State Data
- The validation applies a model trained for 10-minute resting-state scans to pairwise correlations across 65 Seitzman-atlas DMN regions.The resulting denoised estimates are compared with traditional Pearson correlations.Source: §6 · p. 8
6.1 Voxel by Voxel Validation
- Across 411 reliable region pairs, a mean of 46 voxels reached 90% of asymptotic correlation estimates, with requirements varying by pair.The analysis excluded 1,302 of 2,080 pairs that failed the voxel-homogeneity test; that exclusion limits this voxel-reduction analysis, not the method's applicability.Source: §6.1 · pp. 9–10, Figure 2 · p. 11
6.2 Optimizing Scan Duration for Dynamic Connectivity Studies
- A 7-minute 7T scan achieves the same within-session precision as a full 10-minute 3T scan.The reported reduction is approximately 3 minutes, or 28.5% of required scan time.Source: §6.2 · pp. 12–13
- Scan-duration projections rescale posterior standard deviations approximately by 1/√n while preserving posterior shape, a heuristic outside parametric asymptotic settings.For population projections, individual posteriors are rescaled before concatenation because the pooled distribution may be non-Gaussian.Source: §6.2 · p. 13, §6.3 · p. 15
6.3 Across Population Effects
- At the population level, pooled 3T requires roughly 37 times more per-subject scan time than pooled 7T for the curves to converge.The 7T advantage persists and is amplified after pooling across subjects.Source: §6.3 · pp. 14–15
7 Discussion
The discussion presents a Bayesian framework for denoising connectivity estimates and quantifying measurement uncertainty. It relates noise modeling, posterior uncertainty, and voxel averaging to experimental-design choices.
7.1 Summary of Key Findings
- The framework combines coupled Ornstein-Uhlenbeck modeling with amortized neural posterior estimation to denoise connectivity and quantify measurement uncertainty.The discussion frames simulation-based inference as a tool for optimizing scan duration, spatial resolution, and group size for specific scanners and populations.Source: §7.1 · p. 16
7.2 Accounting for Noise Across the Frequency Spectrum
- For |r| ≥0.1 and |r| ≥0.2, reported improvements were 9.92% and 7.82%; 3T gains were 6.9% and negligible, respectively.The framework embeds signal dynamics and residual measurement noise in one generative model for joint separation during Bayesian inference.Source: §7.2 · p. 16, §7.3 · p. 17
7.3 Uncertainty Quantification as a Research Tool
- Posterior coverage was 80% at one standard deviation and 97% at two, versus nominal Gaussian coverage of 68% and 95%.The authors describe these estimates as slightly underconfident, while their width scales appropriately with estimation error.Source: §7.3 · p. 17
- Uncertainty metrics support subject-specific decisions about scan duration, spatial resolution, and acquisition parameters.The discussion contrasts these decisions with universally conservative protocols that may require unnecessarily long scans or coarse parcellations.Source: §7.3 · p. 17
7.4 Voxel-Level Spatial Averaging: Implications for Experimental Design
- Voxel-level validation found that SBI converged with fewer voxels than Pearson correlation, supporting stable estimates from smaller averaging windows.The discussion connects finer parcellations with greater neuroanatomical specificity and less mixing of functionally distinct regions.Source: §7.4 · p. 18
7.5 Optimizing Scan Duration
The section compares scan duration through relative scan length and the time needed for 3T and 7T conclusiveness curves to match. It reports a 7T advantage within session and in an extrapolated large-scan regime, while distinguishing its measure from ICC reproducibility.
7.5 Optimizing Scan Duration
- The analysis uses relative scan length parameter λ and compares the datasets’ thresholds by asking how much more scan time 3T requires than 7T to reach the same conclusiveness level.Source: §7.5 · p. 18
- Within session, 7T used 71.5% of the scan time of 3T for equivalent conclusiveness, meaning 3T required approximately 40% more scan time.The comparison corresponds to approximately 7 minutes at 7T matching a full 10-minute 3T acquisition.Source: §7.5 · p. 18
- These within-session curve intersections are not directly comparable to ICC-based measures of between-session reproducibility.Acquisition parameters and preprocessing can also change the effective information per timepoint.Source: §7.5 · p. 19
- In the extrapolated large-λ regime, 3T required approximately 23% more scan time than 7T to match performance.The passage states that neither threshold is achievable within a practical scanning session, but their ratio indicates that the advantage persists.Source: §7.5 · p. 19
7.6 Across Population Effects
At the population level, 7T requires substantially less per-subject scan time than 3T for equivalent connectivity precision. The comparison is qualified by differences between the datasets and by assumptions in the large-sample extrapolation.
7.6 Across Population Effects
- Within session, pooled 7T matches pooled 3T conclusiveness using 49.2% of the per-subject scan time.This is a roughly 2× improvement over the single-subject local advantage.Source: §7.6 · p. 19
- 37×: pooled 3T requires roughly 37 times more per-subject scan time than pooled 7T to reach the same precision in the extrapolated large-λ regime.The comparison is based on projected curves extended beyond a single session.Source: §7.6 · p. 19
- The datasets differ in population characteristics, acquisition protocols, and preprocessing, so field-strength effects are not isolated.A controlled comparison would ideally scan the same subjects at both field strengths under matched protocols; the extrapolation also relies on 1/√n contraction and Bernstein–von Mises regularity conditions.Source: §7.6 · pp. 19–20
7.7 Methodological Considerations and Future Directions
The paper discusses computational and modeling choices, clinical implications, and the scope of its conclusions. Its comparison finds a substantial 7T population-level scan-time advantage, while noting limits on attributing that advantage to field strength alone.
7.7 Methodological Considerations and Future Directions
- The two-stage inference pipeline is computationally tractable, but surrogate-mapping errors can propagate into denoised correlations and uncertainty bounds.The paper also notes that its seven summary statistics have not been formally optimized and that posterior calibration is conservative.Source: §7.7 · p. 20
- The approach can use alternative generative processes if they can be simulated, while extensions to higher-order dependencies and full network structure remain future directions.The paper also identifies nonlinear dynamics, multiplicative noise, non-stationarity, and region-specific temporal properties as possible model extensions.Source: §7.8 · p. 21
7.8 Clinical Implications
- Clinical deployment still requires controlled field-strength comparisons, multi-site evaluation, and prospective patient studies to test whether improved precision yields better clinical prediction.The framework's generalization across two independent datasets provides initial evidence of robustness.Source: §7.8 · pp. 21–22
7.9 Conclusion
- Pooled 3T requires roughly 37× more per-subject scan time than pooled 7T for equivalent precision.The conclusion attributes the amplified population-level advantage to higher per-subject precision compounding across subjects.Source: §7.9 · p. 22
- Single-subject posterior distributions provide uncertainty quantification without group averaging, a capability the paper identifies as relevant to clinical applications.The conclusion also reports a 7-minute 7T scan matching a full 10-minute 3T acquisition and 46 voxels per ROI achieving 90% of asymptotic precision.Source: §7.9 · p. 22
- A single upfront training investment enables rapid posterior estimation for new observations and supports whole-brain analyses involving thousands of region pairs.The code is provided for deriving scanner-specific and population-specific uncertainty bounds from researchers' own data.Source: §7.9 · p. 22
Data and Code Availability
The paper identifies sources for the framework code and the two datasets.
Data and Code Availability
- The code, trained neural posterior estimator, and decoder network are publicly available on GitHub.The passage gives the repository as https://github.com/Neuroblox/SBI-Correlation-Paper.Source: Data and Code Availability · p. 23
- The HCP Young Adult dataset is available through the Human Connectome Project, while access requests for the Bipolar 7T dataset should go to the corresponding author.The Bipolar dataset was collected as part of the PAgB study.Source: Data and Code Availability · p. 23
Funding
The research was funded by the Baszucki Brain Research Fund, United States.
Funding
- Funding came from the Baszucki Brain Research Fund, United States.The passage identifies LRMP and HHS in connection with the funding.Source: Funding · p. 23