Image harmonization: A review of statistical and deep learning methods for removing batch effects and evaluation...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Data Science

Background:

• Medical imaging data (MRI, CT) from diverse sources exhibit batch effects, confounding downstream analyses.
• These technical variations can bias results and limit the generalizability and reproducibility of neuroimaging studies.
• Existing image harmonization methods aim to mitigate these batch effects.

Purpose of the Study:

• To review and categorize current statistical and deep learning-based image harmonization methods.
• To describe evaluation metrics for assessing harmonization techniques.
• To propose a standardized framework for evaluating new harmonization methods.

Main Methods:

• Systematic review of statistical and deep learning image harmonization approaches.
• Analysis of current metrics for evaluating harmonization effectiveness.
• Development of a standardized evaluation framework.

Main Results:

• Categorization of existing image harmonization methods.
• Identification of key evaluation metrics for harmonization.
• A proposed framework for standardized method evaluation.

Conclusions:

• Effective image harmonization is crucial for reliable neuroimaging research.
• A standardized framework is needed to rigorously evaluate and compare harmonization methods.
• Recommendations are provided for users and methodologists to advance the field.