Scanner invariant representations for diffusion MRI harmonization.

Area of Science:

• Medical Imaging
• Neuroscience
• Data Science

Background:

• Pooled multi-site imaging data exhibit variations due to different sites and scanners.
• Harmonizing imaging data is crucial for large-scale and multi-site studies.
• Existing methods struggle to fully account for inter-scanner variability.

Purpose of the Study:

• To develop and evaluate a novel method for correcting site and scanner biases in diffusion-weighted MRI.
• To create an image reconstruction invariant to original scanning context.
• To improve the reliability of pooled multi-site imaging data.

Main Methods:

• Utilized invariant representation learning, adapted from information theory-based algorithmic fairness.
• Employed a deep learning approach using variational auto-encoders (VAE) to generate scanner-invariant encodings.
• Leveraged the data processing inequality for faithful image reconstruction.

Main Results:

• The proposed invariant representation method demonstrated improvements over a baseline method on independent test data.
• Successfully mapped imaging data between different scanning contexts.
• Achieved image reconstruction that is uninformative of its original source while preserving structural integrity.

Conclusions:

• Invariant representation is a promising technique for harmonizing multi-site diffusion-weighted MRI data.
• The method offers a robust solution for increasing data consistency in large-scale imaging studies.
• This approach supports the growing trend of pooled multi-site imaging research.