Assessing the Reliability of Template-Based Clustering for Tractography in Healthy Human Adults

Jason Kai,Ali R Khan

doi:10.3389/fninf.2022.777853

Abstract

Tractography is a non-invasive technique to investigate the brain’s structural pathways (also referred to as tracts) that connect different brain regions. A commonly used approach for identifying tracts is with template-based clustering, where unsupervised clustering is first performed on a template in order to label corresponding tracts in unseen data. However, the reliability of this approach has not been extensively studied. Here, an investigation into template-based clustering reliability was performed, assessing the output from two datasets: Human Connectome Project (HCP) and MyConnectome project. The effect of intersubject variability on template-based clustering reliability was investigated, as well as the reliability of both deep and superficial white matter tracts. Identified tracts were evaluated by assessing Euclidean distances from a dataset-specific tract average centroid, the volumetric overlap across corresponding tracts, and along-tract agreement of quantitative values. Further, two template-based techniques were employed to evaluate the reliability of different clustering approaches. Reliability assessment can increase the confidence of a tract identifying technique in future applications to study pathways of interest. The two different template-based approaches exhibited similar reliability for identifying both deep white matter tracts and the superficial white matter.

Highlights

The brain consists of numerous regions connected together by axonal bundles which form the structural pathways (Sarwar et al, 2019; Sotiropoulos and Zalesky, 2019) of a highly connected network that enables function and cognition (Mesulam, 1998; Klingberg et al, 1999; Jbabdi et al, 2015; Filley and Fields, 2016)
From the QuickBundle clustered template, an average Euclidean distance of 1.96 ± 0.73 mm and 2.31 ± 0.62 mm was observed for the MyConnectome and Human Connectome Project (HCP) datasets when compared against the average centroid
Distributions of Euclidean distances were similar across datasets, with the MyConnectome dataset exhibiting a lower Euclidean distance against the average centroid for both clustering methods than the HCP dataset

Summary

Introduction

The brain consists of numerous regions connected together by axonal bundles which form the structural pathways ( referred to as tracts) (Sarwar et al, 2019; Sotiropoulos and Zalesky, 2019) of a highly connected network that enables function and cognition (Mesulam, 1998; Klingberg et al, 1999; Jbabdi et al, 2015; Filley and Fields, 2016). The gold standard for investigating structural connectivity are chemical tracers, these techniques are invasive and performed only in animal studies and post-mortem samples (Jbabdi et al, 2015; Sotiropoulos and Zalesky, 2019). Using information from dMRI, an estimation of the pathway trajectories can be reconstructed as a streamline with tractography by (1) estimating the diffusion. An understanding of how tracts are affected in patient cohorts could provide key insights for diagnosis and improve treatment

Methods

Results

Discussion

Conclusion