Sexual Dimorphism in the Corpus Callosum: A Characterization of Local Size Variations and a Classification Driven Approach to Morphometry

David J. Pettey, James C. GeeNeuroImage 2002 · DOI

TL;DR

This paper introduces a novel morphometric approach to analyze local size variations in the corpus callosum (CC) between male and female brains. While statistically significant differences were found, particularly in the splenium, these variations were not substantial enough for reliable classification, highlighting the distinction between statistical significance and diagnostic utility.

What Problem Does This Paper Address?

The study of the corpus callosum (CC), a large white matter tract connecting the two cerebral hemispheres, has revealed potential differences in its structure across various neurological conditions and between sexes. However, previous studies often relied on broad partitions of the CC, potentially missing subtle, localized variations. Furthermore, there's a crucial distinction between identifying statistically significant differences and finding variations that are diagnostically useful for classifying individuals. This paper aims to address these limitations by developing a more granular, quantitative method to characterize local size variations in the CC and to assess their potential for sex-based classification.

Methods

The authors propose a two-pronged quantitative approach:

  1. Statistical Parametric Mapping (SPM) for Local Size Variations:
    • Midsagittal CC slices from male and female subjects are spatially normalized to a common atlas using a non-rigid registration technique (specifically, an elastic-membrane model).
  2. The Jacobian determinant of the deformation field is calculated for each subject. This determinant quantifies local volume changes (expansion or contraction) at each point in the CC relative to the atlas.
  3. To account for overall volume differences, the Jacobian is mean-corrected (normalized by the overall CC area).
  4. The logarithm of the mean-corrected Jacobian is also analyzed, as it's hypothesized to better reflect size changes (e.g., a factor of two increase is a +1 shift in log-space, while a factor of two decrease is a -1 shift).
  5. Pointwise statistics (t-scores and effect sizes) are computed for each "pixel" (or atlas point) to identify statistically significant differences between male and female CCs.
  6. Cluster statistics are then applied to identify contiguous regions of significant differences, using both a classical SPM approach and a permutation test for robustness.
  7. Classification-Driven Approach (Linear Discriminant Function):
    • To assess the diagnostic utility of the identified differences, the authors employ a linear discriminant function (LDF).
  8. This involves performing a principal component analysis (PCA) on the Jacobian fields to reduce the dimensionality of the data and capture the most significant sources of variation.
  9. An LDF is then constructed from these principal components to find the best linear combination of the Jacobian values that discriminates between males and females.
  10. A jack-knife procedure is used to evaluate the classifier's performance by iteratively leaving out one subject and testing the classifier on the remaining data.

Datasets

  • The study utilized a dataset of 29 male and 45 female subjects.
  • One male subject's segmented corpus callosum was used as the atlas for registration.
  • No new benchmarks were introduced; the focus was on applying a novel methodology to existing data.

Key Insights and Ideas

  • Local vs. Global Differences: The study emphasizes that significant differences in the CC might be localized and missed by traditional, broad partitioning methods.
  • Jacobian Determinant as a Measure of Local Size: The Jacobian determinant provides a powerful, continuous measure of local volumetric changes, offering a more nuanced view than simple area measurements.
  • Log-Transformation for Size Changes: The use of the log-transformed Jacobian is theoretically sound for representing size changes symmetrically, addressing potential biases in raw Jacobian values.
  • Statistical Significance vs. Classification Utility: A crucial takeaway is that statistically significant differences (e.g., a low P-value) do not automatically translate into useful diagnostic tools. The LDF analysis revealed that even with significant differences, classification accuracy was only slightly better than chance.
  • Splenium as a Region of Interest: The splenium consistently emerged as a region showing statistically significant differences in relative size between sexes.

Key Results

  • Pointwise Analysis: No statistically significant absolute size differences were found. However, relative (mean-corrected) size differences were observed.
  • SPM Analysis:
    • For the mean-corrected Jacobian ($J_k$), a cluster in the anterior midbody showed a P-value of 0.3.
  • For the log-transformed mean-corrected Jacobian ($j_k^{\text{log}}$), a significant cluster was found in the splenium with a P-value of 0.002.
  • Classification Analysis: The linear discriminant function, even when optimized using principal components and cross-validation, achieved a classification accuracy of only about 60%, indicating that the observed morphological differences, while statistically significant, were not substantial enough for reliable sex classification.

Limitations and Future Work

  • Limited Classification Power: The primary limitation is that the identified statistically significant differences in CC morphology were not sufficient for accurate classification. The authors acknowledge that other measures beyond the Jacobian might yield better classifiers.
  • Assumption of Uniform Correlative Structure: The classical SPM analysis assumes a uniform correlative structure across the CC, which may not hold true. The permutation test offers a more robust alternative by not making this assumption.
  • Future Directions: The authors suggest that their methodology, particularly the classification-driven approach, can be applied to other scalar morphometric quantities and to different populations or conditions. They also hint that exploring different normalization strategies or entirely different morphometric measures might be necessary for achieving better classification performance.

Summary updated

BibTeX
@article{pettey2002sexual,
  author = {David J. Pettey and James C. Gee},
  title = {{Sexual Dimorphism in the Corpus Callosum: A Characterization of Local Size Variations and a Classification Driven Approach to Morphometry}},
  journal = {NeuroImage},
  year = {2002},
  doi = {10.1006/nimg.2002.1130},
}
Figure