Logical circularity in voxel-based analysis: normalization strategy may induce statistical bias.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Statistical Modeling

Background:

• Selection bias is a growing concern in neuroimaging, particularly in voxel-based analyses.
• Template-based registration is crucial for aligning subject images in group studies.
• The choice of similarity metric during registration can impact analysis outcomes.

Purpose of the Study:

• To analytically demonstrate how sum of squared difference (SSD) metrics in template-based registration introduce circularity bias.
• To investigate the influence of different similarity metrics on this bias.
• To propose alternative approaches to mitigate selection bias in voxel-based analyses.

Main Methods:

• Analytical derivation of bias introduced by SSD similarity metrics.
• Hypothetical testing of SSD-related and other similarity metrics (e.g., Demons).
• Evaluation of bias strength based on metric properties.

Main Results:

• Sum of squared difference (SSD) and related metrics explicitly maximize effect size, creating circularity bias.
• The strength of this bias is dependent on the chosen similarity metric.
• Metrics not based on absolute intensity differences show reduced bias.

Conclusions:

• Researchers should exercise caution when using SSD or SSD-related similarity metrics for normalization and statistical analysis within the same image dataset.
• Advocate for using independent image sets for normalization to avoid bias.
• Recommend employing similarity metrics less susceptible to circularity bias.