HACA3: A unified approach for multi-site MR image harmonization.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Magnetic Resonance (MR) imaging suffers from inconsistent acquisition, leading to contrast variations.
• Existing MR harmonization methods often assume identical anatomy across contrasts and require fixed training sets, limiting their flexibility and robustness.
• Sensitivity to imaging artifacts is another significant drawback of current MR harmonization techniques.

Purpose of the Study:

• To introduce Harmonization with Attention-based Contrast, Anatomy, and Artifact Awareness (HACA3), a novel approach for MR image harmonization.
• To address limitations of existing methods, including the questionable assumption of shared anatomy, fixed contrast requirements, and sensitivity to artifacts.
• To develop a versatile and robust MR harmonization technique applicable to diverse datasets and clinical tasks.

Main Methods:

• HACA3 employs an anatomy fusion module to account for inherent anatomical differences across MR contrasts.
• The method is designed for training and application with any combination of MR contrasts, enhancing its flexibility.
• HACA3 incorporates artifact awareness to improve robustness against imaging imperfections.

Main Results:

• HACA3 demonstrates state-of-the-art harmonization performance on diverse MR datasets from 21 sites.
• The approach shows superior performance across multiple image quality metrics compared to existing methods.
• Experiments validate HACA3's versatility and potential clinical impact on tasks like white matter lesion segmentation and longitudinal volumetric analysis.

Conclusions:

• HACA3 effectively harmonizes MR images by accounting for contrast-specific anatomy and artifacts.
• The method offers significant improvements over existing techniques, providing greater flexibility and robustness.
• HACA3 shows promise for enhancing the reliability and applicability of MR imaging in clinical research and practice.