Unpaired T1-weighted MRI synthesis from T2-weighted data using unsupervised learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) provides crucial diagnostic information without ionizing radiation.
• Acquiring multiple MRI sequences (e.g., T1w, T2w) increases scan duration, patient discomfort, and costs.

Purpose of the Study:

• To develop an unsupervised framework for translating T2w MRI images into T1w images.
• To reduce the need for acquiring multiple MRI sequences, thereby streamlining protocols.

Main Methods:

• Utilized a contrast-sensitive domain translation network with adaptive feature normalization.
• Employed adversarial training with cycle consistency, identity, and attention-guided loss functions.
• Ensured preservation of anatomical details and high visual fidelity.

Main Results:

• Achieved a mean PSNR of 22.403 dB, SSIM of 0.775, RMSE of 0.078, and MAE of 0.036.
• Quantitative and qualitative analyses confirmed consistency and perceptual fidelity with ground truth T1w images.
• Demonstrated effective generation of realistic T1w images from T2w inputs.

Conclusions:

• The proposed framework successfully generates clinically acceptable T1w MRI images from T2w inputs.
• This AI-driven approach has the potential to significantly optimize MRI acquisition protocols.
• The method offers a promising solution for reducing scan times and healthcare expenses in medical imaging.