Multi-scanner Harmonization of Paired Neuroimaging Data via Structure Preserving Embedding Learning

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Data Science

Background:

• Combining neuroimaging datasets from multiple sites/scanners increases sample size but introduces technical variability.
• Scanner-related effects can bias downstream analyses, necessitating data harmonization.
• Existing harmonization techniques often struggle with imperfect coregistration and limited scalability.

Purpose of the Study:

• To present MISPEL (Multi-scanner Image harmonization via Structure Preserving Embedding Learning), a novel framework for harmonizing neuroimaging data across multiple scanners.
• To address limitations of existing methods by not assuming perfect coregistration and allowing scalability to more than two scanners.
• To leverage a unique dataset with subjects scanned on four different scanners to define and achieve ideal harmonization.

Main Methods:

• Developed MISPEL, a multi-scanner harmonization framework utilizing Structure Preserving Embedding Learning.
• Incorporated a unique paired dataset where each subject underwent scanning on four distinct scanners.
• Defined an ideal harmonization goal: identical brain tissue volumes for each subject across all scanners.

Main Results:

• MISPEL significantly reduces scanner effects across various metrics.
• The framework demonstrates effectiveness in harmonizing data from multiple scanners.
• The unique dataset enabled precise evaluation of harmonization quality.

Conclusions:

• MISPEL offers a robust solution for harmonizing multi-scanner neuroimaging data.
• The framework improves data consistency, facilitating more accurate pooled analyses.
• MISPEL's ability to handle imperfect coregistration and scale to multiple scanners makes it a valuable tool for large-scale neuroimaging research.