Statistical parametric mapping of brain morphology: Sensitivity is dramatically increased by using brain-extracted images as inputs

George Fein, Bennett A. Landman, Hoang Tran, Jerome Barakos, Kirk Moon, Victoria Di Sclafani, Robert ShumwayNeuroImage 2006 · DOI · OpenReview

TL;DR

This paper demonstrates that removing non-brain tissue (scalp, skull, meninges) from MRI scans before analysis significantly improves the accuracy and sensitivity of voxel-based morphometry (VBM) using Statistical Parametric Mapping (SPM2). This "brain-extracted" approach reduces errors in gray matter segmentation and can potentially halve the required sample size for detecting group differences.

What Problem Does This Paper Address?

Voxel-based morphometry (VBM) is a powerful, automated technique for analyzing structural differences in brain MRI scans, particularly for large datasets. Statistical Parametric Mapping (SPM2) is a popular implementation of VBM. However, the validity and sensitivity of SPM2 have been debated. A key challenge lies in accurately segmenting brain tissue, especially gray matter, from non-brain tissues like the scalp and skull. Errors in this segmentation, often due to misregistration of individual scans with a standard brain template, can lead to incorrect inclusions of non-brain tissue as gray matter or exclusion of actual gray matter. These errors can mask real differences or introduce false positives, reducing the power of VBM analyses. This paper aims to address these segmentation errors and improve the overall sensitivity and reliability of SPM2.

Methods

The core innovation presented is a modified SPM2 processing pipeline that utilizes "brain-extracted" images as input. This involves:

  • Brain Extraction: A crucial preprocessing step where non-brain tissues (scalp, skull, meninges) are removed from the T1-weighted MRI scans. The authors used FSL's Brain Extraction Tool (BET) and an in-house plugin for ImageJ to achieve this.
  • Brain-Extracted Template: A standard MNI152 brain template was also processed to remove non-brain tissue and then smoothed to match the SPM2 pipeline's standard smoothing.
  • Modified SPM2 Pipeline: The standard SPM2 pipeline was adapted to accept these brain-extracted images and the brain-extracted template for registration and analysis.

The authors then evaluated the impact of this modification through:

  • Qualitative Visual Inspection: Manual review of gray matter segmentations for 101 research participants to identify gross and subtle errors.
  • Quantitative Analysis: Calculation of residual mean square errors (RMSEs), which are a direct measure of the statistical model's error and are inversely related to SPM2's sensitivity.

Datasets

The study utilized MRI data from 101 research participants, including abstinent alcoholics and controls. The data was acquired using a 1.5 T GE Signa Infinity scanner. No new benchmarks were introduced; the study focused on improving the analysis of existing MRI data.

Key Insights and Ideas

  • Non-brain tissue is a major source of error: The paper strongly suggests that the inclusion of non-brain tissue in the input images is a primary driver of segmentation errors in SPM2.
  • Misregistration is a critical problem: Errors in aligning individual brain scans to the standard template are not just about including non-brain tissue but also about excluding valid gray matter.
  • Brain extraction dramatically improves segmentation: Removing non-brain tissue before analysis leads to much cleaner and more accurate gray matter segmentations, aligning better with the actual brain boundaries.
  • Sensitivity gains are substantial: The reduction in statistical error (RMSE) directly translates to increased sensitivity, meaning smaller effect sizes can be detected.
  • Sample size reduction is significant: The increased sensitivity implies that researchers may need substantially fewer participants to achieve the same statistical power, potentially halving sample sizes.
  • Sample characteristics matter: Factors like body mass index (BMI) and the presence of scalp fat can exacerbate registration and segmentation errors, highlighting the importance of robust preprocessing.

Key Results

  • Reduced RMSEs: The average RMSE using brain-extracted inputs was less than half that of non-brain-extracted inputs (0.2098 vs. 0.4708). The variance of RMSEs was also significantly lower (one-sixth).
  • Increased Sensitivity: SPM2 analyses using brain-extracted inputs can detect effect sizes about one-third smaller than those using non-brain-extracted inputs.
  • Sample Size Reduction: The authors estimate that sample sizes could be halved for brain-extracted analyses to achieve the same statistical power. For example, an effect size of 0.8 would require 26 subjects per group with non-brain-extracted images but only 12 subjects per group with brain-extracted images.
  • Qualitative Improvement: Visual inspection confirmed a dramatic reduction in both gross and subtle non-brain matter inclusions and improved accuracy in defining gray matter boundaries.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the SPM2's Bayesian segmentation approach, which relies on prior probability templates, might not be universally appropriate, especially for populations with significantly different brain structures (e.g., Alzheimer's patients with larger ventricles).
  • The study focused on T1-weighted images; the applicability to other MRI contrasts is not discussed.
  • While the authors used BET and a custom plugin, the manual removal of residual non-brain tissue suggests that fully automated, robust brain extraction across all datasets remains a challenge.
  • Future Work:
    • The authors suggest exploring segmentation approaches that do not rely on prior probability templates to address potential population-specific biases.
  • Further investigation into how sample characteristics (like BMI) influence registration errors and how to mitigate these effects is implied.
  • The potential for this improved sensitivity to impact other research questions and clinical applications is a natural extension.

Summary updated

BibTeX
@article{fein2006statistical,
  author = {George Fein and Bennett A. Landman and Hoang Tran and Jerome Barakos and Kirk Moon and Victoria Di Sclafani and Robert Shumway},
  title = {{Statistical parametric mapping of brain morphology: Sensitivity is dramatically increased by using brain-extracted images as inputs}},
  journal = {NeuroImage},
  year = {2006},
  doi = {10.1016/j.neuroimage.2005.10.054},
}
Figure