Fast Bayesian whole-brain fMRI analysis with spatial 3D priors

Per Sidén, Anders Eklund, David Bolin, Mattias VillaniNeuroImage 2017 · DOI

TL;DR

This paper introduces a faster, more accurate Bayesian method for analyzing whole-brain fMRI data, addressing the computational challenges of existing approaches. By developing a novel Markov Chain Monte Carlo (MCMC) algorithm and an improved Variational Bayes (VB) method, the authors enable more robust and efficient spatial modeling of brain activity, revealing potential inaccuracies in current standard software.

What Problem Does This Paper Address?

Analyzing functional Magnetic Resonance Imaging (fMRI) data to understand brain activity is computationally intensive. Traditional methods often analyze brain regions separately or use approximations that can lead to inaccuracies. Specifically, the standard Bayesian approach in SPM software (Statistical Parametric Mapping) uses a slice-by-slice analysis with an assumption of independent voxels in the posterior, which the authors argue is a significant limitation. This assumption, while computationally convenient, can lead to spurious findings and an underestimation of true brain activity. The paper aims to develop methods that can perform whole-brain analysis efficiently and accurately, overcoming the computational bottlenecks and the limitations of current approximate methods.

Methods

The authors propose two main methodological contributions:

  • Fast and Practical MCMC: They introduce a Markov Chain Monte Carlo (MCMC) algorithm for exact inference in a spatial Bayesian model. This algorithm is designed to be efficient by exploiting sparsity and using modern techniques for sampling from high-dimensional Gaussian distributions, making whole-brain analysis feasible.
  • Improved Spatial Variational Bayes (SVB): They develop a faster Variational Bayes (VB) algorithm that relaxes the assumption of posterior independence between voxels, which is a key limitation of the existing SPM method (referred to as Independent Variational Bayes or IVB). This new SVB method maintains computational efficiency while providing a more accurate approximation of the true posterior.

Both methods leverage techniques like preconditioned conjugate gradient (PCG) for efficient sampling and solving linear systems, crucial for handling the high dimensionality of fMRI data.

Datasets

The authors evaluate their methods on:

  • Simulated Data: Synthetic fMRI data were generated to assess the scalability and accuracy of the algorithms across different data sizes (10³, 10⁴, and 10⁵ voxels).
  • Real Data:
    • Face Repetition Data: A standard fMRI dataset previously used in Penny et al. (2005b).
  • Object Recognition Data: Data from the OpenfMRI database (Poldrack et al., 2013).

Key Insights and Ideas

  • VB's Achilles' Heel: The assumption of posterior independence between voxels in standard VB methods (like SPM's IVB) can lead to "spurious activation" and a distorted smoothing of brain activity patterns. This is a critical finding, as it suggests current widely-used software might be producing misleading results.
  • MCMC as the Gold Standard (with caveats): The proposed MCMC method provides an "exact" (or at least highly accurate) posterior, serving as a crucial benchmark for evaluating approximate methods. However, its computational cost can still be prohibitive for routine whole-brain analysis.
  • SVB: A Practical Sweet Spot: The improved SVB method offers a compelling balance between speed and accuracy. It significantly outperforms IVB in terms of accuracy and scales much better with data size, making it a viable option for whole-brain analysis.
  • Spatial Priors Matter: The choice of spatial prior (e.g., 2D vs. 3D) has a noticeable impact on the results, influencing shrinkage and the detection of smaller activity regions. The 3D prior generally leads to more conservative estimates.

Key Results

  • Speed: The SVB algorithm is significantly faster than the standard IVB method, especially for large datasets and when run in parallel. For whole-brain analysis with a 3D prior, SVB is generally the fastest option. MCMC is slower but provides the ground truth.
  • Accuracy: SVB demonstrates much higher accuracy compared to IVB, with significantly lower Root Mean Squared Error (RMSE) in posterior mean estimates. IVB shows bias in both mean and standard deviation estimates, particularly for less informative data.
  • Spurious Activations: The paper provides evidence that the factorized VB in SPM12 can lead to spurious activations, a finding not previously highlighted.
  • Scalability: Both SVB and MCMC scale much better with the number of voxels than IVB.

Limitations and Future Work

  • Computational Cost of MCMC: While providing exact inference, MCMC remains computationally demanding for everyday practitioners, especially for whole-brain analysis with many conditions. It is primarily useful as a benchmark.
  • Hyperparameter Estimation: While SVB approximates the posterior well, it can still underestimate hyperparameter variance, and sometimes the mean, especially when data are less informative.
  • Model Selection: The paper highlights the need for robust model selection criteria (e.g., cross-validation, marginal likelihood) to choose between different spatial priors or modeling approaches.
  • Multiple Comparisons: Developing consensus on controlling multiple comparisons in Bayesian spatial models for fMRI remains an open area.
  • Group Analysis: Extending the single-subject analysis to group-level analysis is a significant area for future work, potentially using Bayesian hierarchical models.
  • Prior Flexibility: Exploring more flexible priors, such as Matérn kernels or non-stationary priors, within the efficient PCG sampling framework is a promising direction.
  • GPU Acceleration: Investigating the use of GPUs to further reduce processing time is also suggested.

Summary updated

BibTeX
@article{siden2017fast,
  author = {Per Sidén and Anders Eklund and David Bolin and Mattias Villani},
  title = {{Fast Bayesian whole-brain fMRI analysis with spatial 3D priors}},
  journal = {NeuroImage},
  year = {2017},
  doi = {10.1016/j.neuroimage.2016.11.040},
}
Figure