Wavelet-based Semi-supervised Adversarial Learning for Synthesizing Realistic 7T from 3T MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Ultra-high field 7-tesla (7T) magnetic resonance imaging (MRI) offers superior anatomical detail for diagnosis and prognosis.
• The high cost and inaccessibility of 7T MRI scanners limit their widespread clinical application.

Purpose of the Study:

• To develop a cost-effective method for generating 7T MRI images from readily available 3T scans.
• To address the scarcity of paired 3T-7T MRI data for training deep learning models.

Main Methods:

• A novel wavelet-based semi-supervised adversarial learning framework was proposed.
• The method utilizes a cycle generative adversarial network (CycleGAN) operating in the joint spatial-wavelet domain.
• Leverages unpaired 3T and 7T MRI data to learn the 3T-to-7T image synthesis mapping.

Main Results:

• The proposed semi-supervised method successfully synthesized high-fidelity 7T MR images from 3T counterparts.
• Achieved superior performance compared to state-of-the-art methods trained on fully paired data.
• Demonstrated effective synthesis of multi-frequency details in the joint spatial-wavelet domain.

Conclusions:

• The wavelet-based semi-supervised adversarial learning framework provides a viable solution for generating 7T MRI images.
• This approach enhances the accessibility of advanced 7T MRI imaging by overcoming data limitations.
• The method holds potential for improving diagnostic capabilities in resource-limited settings.