Joint Total Variation ESTATICS for Robust Multi-parameter Mapping

Yaël Balbastre, Mikael Brudfors, Michela Azzarito, Christian Lambert, Martina F. Callaghan, John Ashburner

TL;DR

This paper introduces a new way to clean up high-resolution quantitative MRI (qMRI) data by combining a physics-based signal model with a spatial "Joint Total Variation" prior. By modeling the relationships between different image contrasts and the underlying tissue properties simultaneously, the researchers achieved significantly cleaner parameter maps (like myelin and iron content) without the blurring or bias typical of standard denoising methods.

What Problem Does This Paper Address?

Standard clinical MRI produces "weighted" images where brightness is relative, making it hard to compare scans across different hospitals or time points. Quantitative MRI (qMRI) solves this by measuring absolute physical constants of tissue, such as $R_1$ (longitudinal relaxation) and $R_2^*$ (transverse relaxation), which correlate directly with biological features like myelin and iron content.

However, there is a major catch: to get these measurements at high resolution (e.g., 0.8mm), the signal-to-noise ratio (SNR) drops significantly. The resulting maps are often "noisy" or grainy. While researchers usually fix this by smoothing the images, smoothing destroys fine details and edges. Furthermore, existing methods often treat different image contrasts (like T1-weighted vs. PD-weighted) as independent, even though they are scanning the same brain structure and should share the same underlying edges and anatomy.

Methods

The authors build upon an existing framework called ESTATICS, which performs a joint fit of multiple echoes to estimate $R_2^*$ and signal intercepts. They extend it in two major ways:

  1. Joint Total Variation (JTV) Prior: Instead of denoising each image separately, JTV looks at all image contrasts at once. It assumes that if there is an edge (a change in tissue) in one contrast, there is likely an edge in the others. This "joint" approach allows the algorithm to use information from a high-signal image to help denoise a low-signal one.
  2. Nonlinear Maximum A Posteriori (MAP) Estimation: Rather than just fitting a line to the logarithm of the data (which can be biased by noise), they use a full generative model. They treat the problem as a nonlinear optimization, using clever mathematical tricks like Fisher’s Scoring and Quadratic Bounds to ensure the algorithm converges efficiently to a stable solution.

Notably, the method handles head motion within the model. Instead of "reslicing" images to align them (which can interpolate and corrupt the raw data), they incorporate the motion transformation directly into the signal reconstruction process.

Datasets

The researchers used a "rich" single-subject dataset to validate their approach:

  • MPM Protocol: A single healthy participant was scanned five times in one session using a 0.8 mm Multi-Parameter Mapping protocol.
  • Validation Strategy: Since there is no "perfect" noise-free ground truth in MRI, they used a Leave-One-Echo-Out cross-validation. They trained the model on all but one echo and then tested how accurately the model could "predict" the missing data.

Key Insights and Ideas

  • Edges are Universal: The core "aha" moment is that while $R_1$ and $R_2^*$ represent different physical properties, they share the same structural boundaries (the anatomy of the brain). JTV exploits this structural redundancy to suppress noise while sharpening edges.
  • Physics-Informed Denoising: By embedding the denoising inside the physical signal model (the ESTATICS equations), the algorithm ensures that the resulting "clean" images still obey the laws of MR physics.
  • Validation without Ground Truth: The use of Rice log-likelihood to score "left-out" echoes provides a mathematically rigorous way to prove a denoising method is working without needing a simulated or "perfect" phantom scan.

Key Results

  • Superior Accuracy: In the brain parenchyma (gray and white matter), the JTV method achieved the best predictive score (mean Z-score of 1.19) compared to standard log-linear fits (-0.72) and other state-of-the-art denoising methods like AONLM (-0.41).
  • Vessel Preservation: Qualitatively, the JTV approach was significantly better at preserving small structures like blood vessels, which are often erased by traditional smoothing.
  • Reduced Variance: The method greatly reduced the standard deviation (noise) across the five repeated scans without introducing "bias" (systematic errors in the values). For example, in $R_2^*$ maps, the variance was reduced by nearly half compared to non-regularized methods.

Limitations and Future Work

  • Computational Complexity: The nonlinear MAP estimation is more computationally intensive than simple linear fitting, requiring sophisticated optimization solvers like multi-grid relaxation.
  • Non-Convexity: The forward model is nonlinear, meaning the optimization problem is technically non-convex. While the authors found a unique optimum in practice, theoretical "traps" (local minima) could exist in more complex scenarios.
  • Future Extensions: The authors suggest this framework could be expanded to include super-resolution (reconstructing high-res maps from low-res scans) or incorporating patch-based priors (Non-local means) into the same generative probabilistic framework.

Summary updated

Abstract

Quantitative magnetic resonance imaging (qMRI) derives tissue-specific parameters – such as the apparent transverse relaxation rate $$R_2^\star $$ , the longitudinal relaxation rate $$R_1$$ and the magnetisation transfer saturation – that can be compared across sites and scanners and carry important information about the underlying microstructure. The multi-parameter mapping (MPM) protocol takes advantage of multi-echo acquisitions with variable flip angles to extract these parameters in a clinically acceptable scan time. In this context, ESTATICS performs a joint loglinear fit of multiple echo series to extract $$R_2^\star $$ and multiple extrapolated intercepts, thereby improving robustness to motion and decreasing the variance of the estimators. In this paper, we extend this model in two ways: (1) by introducing a joint total variation (JTV) prior on the intercepts and decay, and (2) by deriving a nonlinear maximum a posteriori estimate. We evaluated the proposed algorithm by predicting left-out echoes in a rich single-subject dataset. In this validation, we outperformed other state-of-the-art methods and additionally showed that the proposed approach greatly reduces the variance of the estimated maps, without introducing bias.
BibTeX
@inproceedings{balbastre2020joint,
  author = {Yaël Balbastre and Mikael Brudfors and Michela Azzarito and Christian Lambert and Martina F. Callaghan and John Ashburner},
  title = {{Joint Total Variation ESTATICS for Robust Multi-parameter Mapping}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  url = {https://openreview.net/forum?id=nYYstWAF_yV},
}

Loading related content...

Figure