Transformation model and constraints cause bias in statistics on deformation fields

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Biomedical engineering

Background:

• Deformation-based morphometry (DBM) relies on nonrigid registration to quantify anatomical differences.
• The choice of registration algorithm and its parameters can influence the accuracy and interpretation of DBM studies.
• Understanding these influences is crucial for reliable neuroimaging research.

Purpose of the Study:

• To investigate the impact of nonrigid transformation models and deformation constraints on DBM study outcomes.
• To compare the performance of popular registration algorithms in terms of deformation field characteristics and Jacobian determinant statistics.

Main Methods:

• Evaluation of three nonrigid registration algorithms: B-spline with constraints, Thirion's demons, and curvature PDE-based.
• Analysis of resulting deformation fields and Jacobian determinant distributions within homogeneous structures.
• Assessment of statistical assumptions, such as gaussianity, for Jacobian values.

Main Results:

• All evaluated algorithms achieved comparable structural overlaps.
• Significant variations were observed in underlying deformation fields and Jacobian determinant values across algorithms.
• Bi-modal Jacobian distributions were found in some regions, violating statistical assumptions.

Conclusions:

• Nonrigid registration algorithm design introduces bias into Jacobian determinant statistics, even with perfect structural overlap.
• These findings have implications for DBM and voxel-based morphometry (VBM) studies employing Jacobian-based modulation.
• Careful consideration of registration methodology is essential for accurate morphometric analyses.