Flexible Amortized Variational Inference in qBOLD MRI

Ivor J. A. Simpson, Ashley McManamon, Alan J. Stone, Nicholas P. Blockley, Alessandro Colasanti, Mara Cercignani

TL;DR

This paper introduces a novel machine learning framework for more accurate and robust estimation of brain oxygen metabolism parameters (OEF and DBV) from qBOLD MRI data. By leveraging synthetic data to create informative priors and employing a flexible variational inference approach, the method overcomes limitations of previous techniques, leading to smoother, more interpretable maps and better identification of physiological changes.

What Problem Does This Paper Address?

Quantitative blood oxygen level dependent (qBOLD) magnetic resonance imaging (MRI) is a powerful tool for mapping brain oxygen metabolism. However, while R2 maps are reliably inferred, estimating the Oxygen Extraction Fraction (OEF) and Deoxygenated Blood Volume (DBV) from qBOLD data is challenging. Existing methods often produce noisy and underestimated OEF maps, alongside overestimation of DBV. This ambiguity stems from the broad joint distribution of probable OEF and DBV values given the MRI signal, making it difficult to disentangle these parameters. This paper aims to address this by developing a more robust and flexible inference framework that can learn to accurately estimate these crucial physiological parameters.

Methods

The core of this work is a flexible amortized variational Bayesian inference framework. The authors propose a two-stage training process:

  1. Synthetic Data Pre-training: A convolutional neural network (encoder) is first trained on synthetically generated qBOLD data. This allows the network to learn informative voxelwise prior distributions for OEF and DBV, based on known ground truth values. This step is crucial for overcoming the inherent ambiguity in real MRI data.
  2. Real Data Fine-tuning: The pre-trained encoder is then fine-tuned on actual MRI data. This stage incorporates a variational inference approach where the network learns to predict the parameters of a scaled multivariate logit-Normal distribution for OEF and DBV. This distribution choice is key as it constrains the parameters to physically plausible ranges and allows for modeling correlations between OEF and DBV. The framework can accommodate any differentiable forward model, and the authors demonstrate the benefit of using the full qBOLD signal equation over asymptotic approximations. Additionally, a total-variation smoothness regularizer is incorporated to encourage spatially smooth parameter maps.

Datasets

  • Training Data: A dataset of 38 scans from 22 subjects was used for training the model.
  • Study Dataset (Validation): A dataset of 6 subjects, with scans acquired at rest and during hyperventilation, was used for validation and analysis.

No new benchmarks were introduced; the study utilizes existing qBOLD MRI data.

Key Insights and Ideas

  • Synthetic Data as Informative Priors: Using synthetic data to pre-train an encoder and derive informative voxelwise priors is a powerful strategy to overcome the ill-posed nature of inferring OEF and DBV from qBOLD MRI. This allows the model to learn from known ground truth and guide inference on real data.
  • Logit-Normal Distribution for Constrained Inference: Modeling OEF and DBV jointly with a scaled multivariate logit-Normal distribution is crucial. It not only constrains the parameters to physically realistic ranges (e.g., OEF between 0% and 100%) but also allows for capturing potential correlations between them, leading to more accurate and interpretable results.
  • Full Forward Model Advantage: Employing the full qBOLD signal equation, rather than asymptotic approximations, significantly improves the model's ability to explain the data, as evidenced by higher ELBO values.
  • Spatial Smoothness as a Regularizer: Incorporating a total-variation smoothness loss is beneficial for producing more interpretable OEF and DBV maps, especially in noisy qBOLD data, by leveraging the biological expectation of spatial continuity.

Key Results

  • The proposed method produces smoother and more interpretable OEF and DBV maps compared to traditional weighted least squares (WLS) methods.
  • The framework demonstrates a substantial improvement in localizing differences in OEF between conditions (e.g., baseline vs. hyperventilation) compared to WLS.
  • The model shows significant increases in OEF and R2 during hyperventilation, consistent with physiological expectations.
  • The full forward model with two compartments yielded the highest ELBO, indicating better data explanation.
  • The use of a full covariance matrix for OEF and DBV (modeling their correlation) led to improved ELBO compared to a diagonal matrix.

Limitations and Future Work

  • Bias from Priors: The authors acknowledge that the implied bias of the prior may dampen any inferred deviations, suggesting that further investigation into the influence of prior distributions, especially in disease contexts, is needed.
  • Susceptibility Artifacts: The method may still be affected by susceptibility artifacts, particularly in deep gray matter regions or in cases of neuroinflammation.
  • Future Work:
    • Integrating cerebral blood flow estimation from arterial spin labeling (ASL) data for more comprehensive inference.
  • Developing more principled methods for integrating spatial smoothness priors.
  • Experimenting with different compartment models.
  • Investigating the impact of tissue-specific priors, especially for challenging regions.

Summary updated

Abstract

Streamlined qBOLD acquisitions enable experimentally straightforward observations of brain oxygen metabolism. $R_2^\prime$ maps are easily inferred; however, the Oxygen extraction fraction (OEF) and deoxygenated blood volume (DBV) are more ambiguously determined from the data. As such, existing inference methods tend to yield very noisy and underestimated OEF maps, while overestimating DBV. This work describes a novel probabilistic machine learning approach that can infer plausible distributions of OEF and DBV. Initially, we create a model that produces informative voxelwise prior distribution based on synthetic training data. Contrary to prior work, we model the joint distribution of OEF and DBV through a scaled multivariate logit-Normal distribution, which enables the values to be constrained within a plausible range. The prior distribution model is used to train an efficient amortized variational Bayesian inference model. This model learns to infer OEF and DBV by predicting real image data, with few training data required, using the signal equations as a forward model. We demonstrate that our approach enables the inference of smooth OEF and DBV maps, with a physiologically plausible distribution that can be adapted through specification of an informative prior distribution. Other benefits include model comparison (via the evidence lower bound) and uncertainty quantification for identifying image artefacts. Results are demonstrated on a small study comparing subjects undergoing hyperventilation and at rest. We illustrate that the proposed approach allows measurement of gray matter differences in OEF and DBV and enables voxelwise comparison between conditions, where we observe significant increases in OEF and $R_2^\prime$ during hyperventilation.
BibTeX
@misc{simpson2022flexible,
  author = {Ivor J. A. Simpson and Ashley McManamon and Alan J. Stone and Nicholas P. Blockley and Alessandro Colasanti and Mara Cercignani},
  title = {{Flexible Amortized Variational Inference in qBOLD MRI}},
  year = {2022},
  eprint = {2203.05845},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2203.05845},
}

Loading related content...

Figure