Image dataset harmonization

Last revised by Candace Makeda Moore on 27 Feb 2024

Image dataset harmonization is the process of changing groups of images acquired with different acquisition parameters and/or in different machines, or their derived values to be as if acquisition was similar. Differences in derived values for characteristics of radiological images in images from differently acquired datasets may reflect only differences in machines and acquisitions instead of underlying biological differences in subjects. For example, if an extremely high resolution MRI image of the same brain is compared to a lower resolution one, what were read as two adjacent white matter hyperintensities may appear as one.  Because large datasets for machine learning are often acquired at different institutions on different machines, harmonization is often critical to improve machine learning outcomes.

The classic, and still broadly used algorithms for harmonization of radiology images, are derivatives of and improvements on the Combat algorithm e.g. neurocombat 1, Combat ++ 2, et cetera. Newer approaches to image and derived dataset harmonization include the use of generative AI algorithms.

ADVERTISEMENT: Supporters see fewer/no ads

Updating… Please wait.

 Unable to process the form. Check for errors and try again.

 Thank you for updating your details.