Abstract
Traditional morphometric analysis often relies on Principal Component Analysis (PCA) for dimensionality reduction, which frequently results in global, non-specific components that are difficult to interpret biologically. This paper, presented at MICCAI 1999, proposes the use of Exploratory Factor Analysis (EFA) to identify localized, anatomically significant clusters in high-dimensional brain imaging data. By analyzing Jacobian determinant fields from registered MRI scans of the corpus callosum, the method successfully isolates regions driving gender-based morphological differences without prior manual labeling.
Introduction: Beyond Variance-Based Compression
In computational anatomy, the goal is often to represent shape variability in a way that generates clinical hypotheses. While PCA is a standard tool for data reduction, its principal components typically represent a mixture of changes across an entire organ. This lack of localization makes biological interpretation challenging.
The authors address this by seeking a method that identifies localized clusters of variables that covary. They propose Factor Analysis (FA) combined with Varimax rotation to discover "factors" of anatomical change in datasets where the number of voxels significantly exceeds the number of subjects.
Methodology
The proposed pipeline integrates non-rigid image registration with a refined statistical framework:
- Non-Rigid Registration: MRI scans are elastically deformed to a template. The local expansion or contraction is captured by the Jacobian determinant at each voxel.
- Factor Analysis (FA): Unlike PCA, which maximizes variance, FA models observed variables as linear combinations of underlying "common factors." This specifically targets the shared behavior (covariance) between anatomical regions.
- Varimax Rotation: A Varimax rotation is applied to the loading matrix to encourage sparsity. This ensures that each anatomical point is primarily associated with a single factor, enhancing interpretability.
- Iterative Pruning: An algorithm automatically determines the optimal number of factors by starting with a large set and iteratively discarding those that do not show a strong correlation (absolute value $\ge 0.5$) with at least two variables.
Key Results and Biological Insights
The study applied this method to a dataset of 28 MRI scans (12 male, 16 female) focusing on the midsagittal corpus callosum.
Unsupervised Discovery of Sub-regions
The iterative algorithm reduced an initial 27 factors to 18 meaningful ones. A key finding was that the splenium (the posterior part of the corpus callosum) is not a uniform structure. Factor analysis revealed distinct sub-regions that behave differently across genders.
Gender Dimorphism
The analysis pinpointed a significant difference in the left region of the inferior splenium, which was found to be larger in females. In contrast, the right half of the same structure showed no significant gender difference. One specific factor (Factor 15) isolated a region in the right-most splenium with a high effect size (> 0.75), highlighting a clear target for further biological study.
Discussion and Future Directions
The EFA approach demonstrates that preserving correlation between variables is more valuable for anatomical discovery than simply maximizing variance. While the current study was limited by a small sample size (n=28) and a 2D focus, it provides a robust framework for unsupervised morphometry.
Future work includes extending the method to 3D volumetric data and applying it to clinical populations, such as identifying early sub-regional atrophy in Alzheimer's disease. The use of heuristic thresholds for factor pruning remains a point for refinement in subsequent iterations of the model.
Summary updated