Power estimation for non-standardized multisite studies

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Statistical Modeling

Background:

• Multisite neuroimaging studies face challenges with scanner and sequence variability potentially obscuring true biological effects.
• Current harmonization methods require standardization or phantom-based corrections, which can be resource-intensive and impractical.

Purpose of the Study:

• To propose a novel statistical framework that bypasses the need for data harmonization and phantom-based corrections in multisite neuroimaging studies.
• To develop a power equation for defining site inclusion criteria based on the variability of estimated regional volume scaling factors.

Main Methods:

• Estimated scaling factors for 20 heterogeneous scanners across the US and Europe using a single cohort of 12 subjects.
• Applied Freesurfer's segmentation algorithm for regional volume estimation and ordinary least squares for scaling factor calculation.
• Validated scaling factors through power curve comparisons, leave-one-out calibration, and pre/post-calibration agreement analysis.

Main Results:

• Demonstrated that regional volume bias scales between sites due to scanner and sequence differences.
• Derived a power equation enabling the determination of conditions where harmonization is unnecessary to achieve 80% statistical power.
• Successfully defined inclusion criteria for multisite studies based on scaling factor variability.

Conclusions:

• The proposed framework offers a method to inform the selection of processing pipelines and outcome metrics for multisite studies.
• This approach facilitates collaborations between diverse clinical and research institutions by providing a data-driven site selection strategy.
• Enables robust multisite neuroimaging research without the need for extensive data harmonization or phantom calibration.