Image Synthesis in Multi-Contrast MRI With Conditional Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-contrast magnetic resonance imaging (MRI) enhances diagnostic information but is limited by scan time and potential artifacts.
• Current multi-contrast synthesis methods often lose structural details due to nonlinear intensity transformations.
• Synthesizing unacquired or corrupted contrasts can significantly improve MRI's diagnostic utility.

Purpose of the Study:

• To develop a novel approach for multi-contrast MRI synthesis using conditional generative adversarial networks (cGANs).
• To preserve high-frequency details and improve synthesis performance for both registered and unregistered multi-contrast images.
• To leverage information from neighboring cross-sections for enhanced synthesis quality.

Main Methods:

• Implementation of a conditional generative adversarial network (cGAN) framework for MRI synthesis.
• Integration of adversarial, pixel-wise, perceptual, and cycle-consistency losses to enhance synthesis fidelity.
• Utilization of cross-sectional information from adjacent image slices to improve detail preservation.

Main Results:

• The proposed cGAN approach demonstrated superior performance in synthesizing multi-contrast MRI compared to existing methods.
• Preservation of intermediate-to-high frequency details was achieved through the adversarial loss.
• Enhanced synthesis quality was observed on T1- and T2-weighted images from both healthy subjects and patients.

Conclusions:

• The novel cGAN-based synthesis approach significantly improves the quality and versatility of multi-contrast MRI exams.
• This method offers a solution to limitations imposed by scan time and image artifacts in MRI acquisition.
• The technique can potentially reduce the need for prolonged or repeated MRI examinations, improving patient experience and diagnostic efficiency.