An empirical comparison of surface-based and volume-based group studies in neuroimaging

Alan Tucholka, Virgile Fritsch, Jean-Baptiste Poline, Bertrand Thirion

TL;DR

This paper empirically compares two main approaches for analyzing fMRI group data: surface-based (2D) and volume-based (3D). They found that surface-based methods generally offer greater sensitivity and reproducibility in detecting brain activity, suggesting they may be a more powerful tool for neuroimaging research.

What Problem Does This Paper Address?

Neuroimaging studies, particularly those using fMRI, aim to pinpoint where in the brain specific cognitive functions occur. A common challenge is that brain structures vary significantly between individuals. To overcome this, researchers typically "normalize" individual brains to a standard template, allowing for group-level analysis. Traditionally, this normalization and subsequent analysis are done in 3D space (volume-based). However, most fMRI activity is localized to the brain's cortex, which has a complex, folded surface. Analyzing this surface directly (surface-based) might offer better alignment and thus more sensitive detection of activity. Despite the theoretical advantages of surface-based methods, there's been a lack of comprehensive empirical comparisons with standard volume-based approaches, especially concerning group-level statistical inference. This paper aims to fill that gap by directly comparing the performance of these two approaches.

Methods

The authors employed a variety of statistical methods to compare surface-based and volume-based analyses. They used both Random Effects (RFX) and Mixed Effects (MFX) models, which account for variability both within and between subjects. For detecting activation, they used both voxel-level (or vertex-level for surface) and cluster-level statistics.

To compare the results, they developed three main strategies:

  1. Projection Approach: They projected the 3D volume-based results onto the cortical surface of each individual and compared the extent of activated areas.
  2. Embedding Approach: They embedded the surface-based results into the 3D MNI space and compared their spatial similarity to the volume-based results using a kernel-based metric.
  3. Reproducibility Analysis: They used a bootstrap procedure to assess how reliably active regions were detected across different subsets of subjects.

They also investigated the impact of smoothing (applying a spatial filter) at different levels for both surface and volume analyses.

Datasets

The study utilized fMRI data from 25 healthy subjects who underwent a functional localizer protocol designed to activate various brain regions. This protocol included tasks like button presses, mental computation, and sentence processing. The data was acquired using a 1.5T GE scanner. No new benchmark datasets were introduced; the focus was on analyzing existing fMRI data with different methods.

Key Insights and Ideas

  • Surface-based analysis is generally more sensitive: This means it tends to detect more activated regions, especially when looking at individual activation peaks.
  • Better inter-subject correspondence on the surface: The complex folding of the cortex, when properly aligned across individuals, seems to reduce spatial uncertainty and improve the detection of subtle activations.
  • Cluster-level inference in volume can be misleading: In some cases, volume-based cluster analysis can merge spatially distant activations from different subjects into a single large cluster, potentially misrepresenting the underlying functional organization. Surface-based analysis, by contrast, can better delineate these distinct regions.
  • Reproducibility is higher on the surface: Surface-based methods consistently showed more reproducible results, suggesting they are more robust to individual variability.
  • Smoothing impacts both, but surface still leads: While smoothing generally increases the size and reproducibility of detected regions in both approaches, surface-based methods still maintain an advantage.

Key Results

  • Increased Activated Area: Surface-based analysis consistently showed a larger relative volume (in percent) of activated areas compared to volume-based analysis across various contrasts and statistical tests (e.g., CMFX surface showed 9.221% for computation-sentences vs. 2.075% for volume).
  • Higher Reproducibility: Surface-based methods achieved higher reproducibility indices (around 0.7-0.8) compared to volume-based methods (around 0.4-0.6) across different smoothing levels.
  • Spatial Specificity: The embedding approach revealed that surface-based peak-level inference could reveal more detailed activation patterns, while volume-based cluster inference tended to produce larger, less specific regions.
  • Kernel-based Similarity: The similarity index between surface and volume results was generally high (around 0.8), indicating a good overall correspondence, but differences emerged, particularly in peak-level inference where surface-based methods seemed to capture more fine-grained details.

Limitations and Future Work

The authors acknowledge several limitations:

  • Projection Accuracy: The accuracy of projecting fMRI data onto the cortical surface is a critical step and can be affected by factors like fMRI resolution, EPI distortion correction, and smoothing. This step is identified as the riskiest for surface-based inference.
  • Pipeline Dependence: The results might be influenced by the specific processing pipelines and tools used.
  • Limited Scope: They did not conduct a systematic study of different pipelines or the impact of stronger fMRI contrasts.

Future work could focus on:

  • Developing quality check measures for the accuracy of fMRI signal projection onto the cortical surface.
  • Investigating the impact of different processing pipelines and parameters.
  • Exploring how stronger fMRI contrasts might further influence the comparison between surface and volume-based analyses.
  • Comparing surface co-registration based on geometric features (like FreeSurfer) with other surface representations.

Summary updated

Abstract

Being able to detect reliably functional activity in a population of subjects is crucial in human brain mapping, both for the understanding of cognitive functions in normal subjects and for the analysis of patient data. The usual approach proceeds by normalizing brain volumes to a common three-dimensional template. However, a large part of the data acquired in fMRI aims at localizing cortical activity, and methods working on the cortical surface may provide better inter-subject registration than the standard procedures that process the data in the volume. Nevertheless, few assessments of the performance of surface-based (2D) versus volume-based (3D) procedures have been shown so far, mostly because inter-subject cortical surface maps are not easily obtained. In this paper we present a systematic comparison of 2D versus 3D group-level inference procedures, by using cluster-level and voxel-level statistics assessed by permutation, in random effects (RFX) and mixed-effects analyses (MFX). We consider different schemes to perform meaningful comparisons between thresholded statistical maps in the volume and on the cortical surface. We find that surface-based multi-subject statistical analyses are generally more sensitive than their volume-based counterpart, in the sense that they detect slightly denser networks of regions when performing peak-level detection; this effect is less clear for cluster-level inference and is reduced by smoothing. Surface-based inference also increases the reliability of the activation maps. Previous article in issue Next article in issue
BibTeX
@article{tucholka2012empirical,
  author = {Alan Tucholka and Virgile Fritsch and Jean-Baptiste Poline and Bertrand Thirion},
  title = {{An empirical comparison of surface-based and volume-based group studies in neuroimaging}},
  journal = {NeuroImage},
  year = {2012},
  doi = {10.1016/j.neuroimage.2012.06.019},
}
Figure