Acquisition-invariant brain MRI segmentation with informative uncertainties.

Area of Science:

• Medical Imaging Analysis
• Data Harmonization
• Machine Learning

Background:

• Combining multi-site data offers benefits but is challenged by site-specific covariates that bias analyses.
• Existing post-hoc correction methods often rely on unmet assumptions in real-world applications.

Purpose of the Study:

• To develop an algorithm robust to acquisition physics and site-specific effects in segmentation tasks.
• To incorporate explicit uncertainty modeling for identifying generalization failures.

Main Methods:

• An algorithm designed to account for site-specific effects, including sequence parameter choices.
• Integration of uncertainty modeling to detect algorithm generalization failures.
• Demonstration on segmentation tasks within medical imaging.

Main Results:

• The proposed method generalizes to holdout datasets, maintaining segmentation quality.
• The algorithm successfully accounts for site-specific sequence choices, acting as a harmonization tool.
• Explicit uncertainty modeling identifies potential generalization failures.

Conclusions:

• The developed algorithm provides a robust approach to multi-site data harmonization in medical imaging.
• The method enhances data reliability by accounting for acquisition variability and modeling uncertainty.
• This work facilitates more accurate and generalizable downstream analyses from combined datasets.