A Continuous Model of Cortical Connectivity

Daniel Moyer, Boris A. Gutman, Joshua Faskowitz, Neda Jahanshad, Paul M. Thompson

TL;DR

This paper introduces a novel continuous model for brain connectivity using a Poisson point process, treating individual white matter tracts as events in a continuous space. This approach offers a more robust and reliable representation of structural connectivity compared to traditional discrete graph-based methods, showing significantly higher test-retest reliability.

What Problem Does This Paper Address?

The study of brain connectivity, particularly structural connectivity derived from diffusion MRI, has largely relied on representing the brain as a network of discrete regions (nodes) connected by structural pathways (edges). This "region-to-region" paradigm is heavily dependent on how the brain is parcellated into regions. Different parcellations can lead to vastly different connectivity statistics, making it difficult to compare findings across studies or to definitively determine the "optimal" way to divide the brain. This paper addresses the need for a more continuous and parcellation-independent framework for characterizing brain connectivity, aiming to provide a more stable and reliable measure of structural connections.

Methods

The core of the paper is a continuous model of cortical connectivity based on a Poisson point process. This model treats each observed white matter tract (streamline) as an "event" in a continuous "connectome space," which is defined as the product space of cortical white matter boundaries. The intensity function of this Poisson process, which describes the probability of finding a tract endpoint pair at a given location, is estimated using non-parametric Kernel Density Estimation (KDE).

To handle the computational burden of KDE on a high-dimensional space, the authors leverage the properties of the spherical heat kernel and its truncated spherical harmonic representation. This allows for a fast parameter estimation method by pre-computing associated Legendre products. The model is then used to derive region-to-region connectivity matrices for any given parcellation. The paper also proposes two measures to evaluate the quality of a parcellation: Approximation Error (measuring how well the discrete graph approximates the continuous intensity function) and Likelihood (leveraging the point process model's probabilistic nature).

Datasets

The study uses data from the Consortium for Reliability and Reproducibility (CoRR) dataset, specifically a sub-dataset of 29 subjects from the Institute of Psychology, Chinese Academy of Sciences. Each subject underwent two diffusion-weighted MRI (DWI) scans approximately two weeks apart, allowing for test-retest reliability assessments. The paper also mentions using three standard cortical parcellations: the Destrieux atlas, the Desikan-Killiany (DK) atlas, and the Desikan-Killiany-Tourville (DKT31) atlas.

Key Insights and Ideas

  • Continuity as a Superior Representation: By modeling connectivity as a continuous intensity function rather than discrete connections, the approach inherently smooths over noise and minor variations in tractography, leading to more stable and reliable estimates.
  • Parcellation Independence: The continuous model provides a foundational representation of connectivity that is independent of any specific brain parcellation. This allows for a more objective comparison of different parcellation schemes.
  • Leveraging Spherical Harmonics for Efficiency: The use of spherical harmonics for the heat kernel significantly speeds up the estimation process, making the continuous model computationally feasible.
  • Reliability Matters: The paper strongly suggests that continuous connectomes are substantially more reliable (higher test-retest ICC scores) than traditional discrete count-based connectomes, especially when dealing with noisy data or small but non-zero connections.

Key Results

  • The proposed integrated intensity connectomes achieved significantly higher Intraclass Correlation (ICC) scores for test-retest reliability compared to standard fiber endpoint count methods. For example, with thresholding, the intensity method achieved ICC scores of 0.5613 (DK), 0.4868 (Destrieux), and 0.4388 (DKT31), while the count method achieved 0.2093, 0.1722, and 0.2266 respectively.
  • The paper demonstrates that the continuous model can be used to assess the quality of different cortical parcellations using measures like ISE, Negative Log Likelihood, and AIC scores.
  • While all three parcellations showed consistent results across measures, the Destrieux atlas, despite having more regions, showed a lower likelihood criterion, possibly due to the increased number of region pairs.

Limitations and Future Work

  • Computational Cost: While the paper presents a fast estimation method, the initial computation of Legendre products and the overall process can still be computationally intensive.
  • Tractography Method Dependence: The results are conditioned on the use of a specific probabilistic CSD tractography model. Different tractography methods might yield different intensity functions and resulting matrices.
  • Future Work: The authors suggest exploring other adaptations, such as using FA-weighted fiber counting, conducting tests on larger datasets, and investigating the relative differences between various tracking methods and parcellations more rigorously. They also acknowledge that the choice of gray matter dilation for fiber counting versus streamline projection remains an open question.

Summary updated

Abstract

We present a continuous model for structural brain connectivity based on the Poisson point process. The model treats each streamline curve in a tractography as an observed event in connectome space, here a product space of cortical white matter boundaries. We approximate the model parameter via kernel density estimation. To deal with the heavy computational burden, we develop a fast parameter estimation method by pre-computing associated Legendre products of the data, leveraging properties of the spherical heat kernel. We show how our approach can be used to assess the quality of cortical parcellations with respect to connectivty. We further present empirical results that suggest the “discrete” connectomes derived from our model have substantially higher test-retest reliability compared to standard methods.
BibTeX
@inproceedings{moyer2016continuous,
  author = {Daniel Moyer and Boris A. Gutman and Joshua Faskowitz and Neda Jahanshad and Paul M. Thompson},
  title = {{A Continuous Model of Cortical Connectivity}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2016},
  url = {https://openreview.net/forum?id=E_qd-xHXUTH},
}
Figure