A biophysical model of brain deformation to simulate and analyze longitudinal MRIs of patients with Alzheimer's disease

Bishesh Khanal, Marco Lorenzi, Nicholas Ayache, Xavier PennecNeuroImage 2016 · DOI

TL;DR

This paper introduces a novel biophysical model for simulating brain deformation in Alzheimer's Disease (AD) patients. By accounting for the distinct mechanical properties of brain tissue and cerebrospinal fluid (CSF), the model can generate realistic longitudinal MRI sequences, offering a powerful tool for understanding AD progression and evaluating atrophy estimation algorithms.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) causes progressive brain atrophy, which is observable in longitudinal MRI scans. Accurately simulating these changes is crucial for understanding disease mechanisms, developing treatments, and evaluating the algorithms used to detect and quantify atrophy. Existing simulation methods often oversimplify the complex biomechanics of brain tissue and the compensatory expansion of CSF, leading to unrealistic deformations and limiting their utility for rigorous algorithm benchmarking. This paper aims to bridge this gap by proposing a more physically grounded model that captures these crucial aspects.

Methods

The core of this work is a biophysical model for brain deformation that treats brain parenchyma and CSF distinctly. The model is based on continuum mechanics and incorporates a tissue loss model that accounts for mass and volume reduction while maintaining constant density. Key aspects include:

  • Conservation Law: A modified conservation law is derived to account for mass loss due to atrophy, expressed as $\nabla \cdot u = -\alpha$, where $\alpha$ represents the local volume loss rate.
  • Elastic Energy Minimization: The deformation is modeled as minimizing strain energy, using a Saint Venant-Kirchoff model. This leads to a system of partial differential equations (PDEs) similar to Stokes flow equations, but with a non-zero mass source term.
  • CSF Modeling: The CSF region is modeled with a different set of equations that allow for expansion to compensate for tissue volume loss, with a compressibility coefficient $k$.
  • Numerical Implementation: The PDEs are solved numerically using a staggered grid discretization and the Finite Difference Method (FDM), chosen for its stability and efficiency compared to Finite Element Method (FEM).
  • Pipeline for Simulation: A comprehensive pipeline is presented for simulating longitudinal MRIs, starting from a baseline MRI, requiring skull stripping and segmentation into brain parenchyma, CSF, and skull regions. An atrophy map is then prescribed, and the model generates the corresponding deformation field and simulated follow-up image.

Datasets

The paper utilizes the MIRIAD dataset, which contains T1 structural MRIs of 46 Alzheimer's patients with multiple time-point images spanning up to 2 years. This dataset is used for experimental validation and to create simulated follow-up images for evaluating atrophy estimation algorithms.

Key Insights and Ideas

  • Differential Tissue Properties Matter: Treating brain tissue and CSF with different biomechanical properties is crucial for realistic simulation. CSF expansion plays a vital role in compensating for tissue volume loss, and its behavior is distinct from brain tissue.
  • Biophysical Constraints Improve Realism: The model's foundation in biomechanical principles allows for more plausible deformations compared to purely geometric or Jacobian-based methods.
  • Atrophy Estimation Algorithm Evaluation: The developed framework provides a powerful tool for quantitatively assessing the performance and biases of atrophy estimation algorithms under controlled conditions, by simulating ground truth data.
  • Model Parameters Influence Deformation: Even with the same prescribed atrophy map, varying biomechanical parameters like shear modulus ($\mu$) and compressibility ($k$) can lead to different deformation fields, highlighting the importance of accurate parameter estimation.

Key Results

  • The model successfully simulates realistic brain deformations, including sulcal widening and ventricular enlargement, consistent with observed AD progression.
  • Experiments demonstrate that different spatial distributions of atrophy, even with the same global volume loss, can lead to similar simulated images, posing a challenge for atrophy estimation algorithms that rely solely on image appearance.
  • The framework allows for the simulation of complex, patient-specific atrophy patterns, enabling a more thorough evaluation of atrophy estimation algorithms.
  • When comparing atrophy estimation from real vs. simulated follow-up images, the study found reduced variability in estimates when using simulated data, suggesting that the controlled nature of simulations can help isolate sources of error.

Limitations and Future Work

  • Computational Cost: Solving the model is computationally intensive, requiring significant processing time even with distributed computing. This poses a challenge for optimization frameworks that involve numerous parameter estimations.
  • Parameter Calibration: While the model has biomechanically inspired parameters, their precise values for specific brain regions and disease stages are not fully established and require further investigation and calibration.
  • Integration of Other Modalities: The authors suggest that incorporating information from other imaging modalities (e.g., diffusion imaging for structural connectivity, PET for tau pathology) could enhance the accuracy of the "Atrophy Generation" block.
  • Anisotropy: The anisotropic nature of brain parenchyma due to fibers is not explicitly modeled but is identified as an interesting avenue for future exploration.
  • Noise and Artifacts: While the model focuses on deformation, the impact of image noise and acquisition artifacts on atrophy estimation is acknowledged but not fully addressed in this work.

Summary updated

BibTeX
@article{khanal2016biophysical,
  author = {Bishesh Khanal and Marco Lorenzi and Nicholas Ayache and Xavier Pennec},
  title = {{A biophysical model of brain deformation to simulate and analyze longitudinal MRIs of patients with Alzheimer's disease}},
  journal = {NeuroImage},
  year = {2016},
  doi = {10.1016/j.neuroimage.2016.03.061},
}

Loading related content...

Figure