Harmonizing three-dimensional MRI using pseudo-warping field guided GAN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automated diagnostic models for magnetic resonance imaging (MRI) require large, multisite, and heterogeneous datasets.
• Direct application of such data can lead to biased outcomes.
• Existing MRI harmonization methods often focus on 2D slices, causing inter-slice variations in 3D volumes.

Purpose of the Study:

• To develop a novel method for harmonizing MRI data that effectively addresses inter-slice inconsistencies.
• To improve the reliability and accuracy of automated diagnostic models by reducing data heterogeneity.
• To enhance the spatial accuracy and slice-wise consistency of 3D MRI volumes.

Main Methods:

• Introduction of a pseudo-warping field to artificially transform MRI slices, ensuring consistent harmonization across adjacent slices.
• Development of unsupervised spatial and recycle loss functions to enforce spatial accuracy and slice-wise consistency.
• Utilizing a generative approach that operates on slices rather than full 3D volumes.

Main Results:

• The proposed model successfully mitigates inter-slice variations in MRI data.
• Anatomical details within the images are effectively preserved during the harmonization process.
• The model demonstrates superior computational efficiency and flexibility compared to existing 3D generative harmonization methods.

Conclusions:

• The pseudo-warping field approach offers an effective solution for harmonizing 3D MRI data by addressing inter-slice variations.
• This method enhances the quality of imaging datasets for developing robust automated diagnostic models.
• The model's efficiency and flexibility make it a promising tool for neuroimaging research and clinical applications.