Style transfer generative adversarial networks to harmonize multisite MRI to a single reference image to avoid...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Multisite neuroimaging studies require harmonizing magnetic resonance (MR) images acquired across different scanners and protocols.
• Existing retrospective harmonization methods may over-correct for technical variations, confounding them with population variability.
• Current statistical approaches often necessitate similar demographic or clinical data across datasets to isolate acquisition-based variability.

Purpose of the Study:

• To develop a novel method for harmonizing multisite MR images by treating it as a style transfer problem.
• To overcome limitations of existing methods that cannot distinguish technical variations from population variability.
• To enable high-powered neuroimaging analyses by improving data consistency across diverse datasets.

Main Methods:

• A fully unsupervised deep-learning framework based on a generative adversarial network (GAN) was employed.
• The model harmonizes MR images by encoding and inserting style information from a reference image, independent of site/scanner labels.
• The model was trained on data from five large-scale, multisite datasets with varied demographics.

Main Results:

• The style-encoding GAN model successfully harmonized MR images and matched intensity profiles without relying on traveling subjects or site labels.
• The method effectively removed site-related variances while preserving crucial anatomical information and clinically meaningful patterns.
• Harmonization improved extracted features, brain-age estimates, and case-control effect sizes, demonstrating its clinical utility.
• The model generalized to unseen scanners and protocols when trained on a diverse dataset.

Conclusions:

• The proposed style transfer approach offers an effective solution for MR image harmonization in multisite neuroimaging studies.
• This method enhances data consistency and comparability, facilitating more robust analyses of psychiatric and neurological conditions.
• The unsupervised, label-agnostic framework provides a promising tool for collaborative research, overcoming common data heterogeneity challenges.