Self-supervised adversarial diffusion models for fast MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Deep Learning for Medical Image Reconstruction

Background:

• Magnetic resonance imaging (MRI) provides crucial soft tissue contrast for medical diagnosis and treatment.
• However, lengthy MRI acquisition times lead to patient discomfort and motion artifacts, compromising image quality.
• Accelerating MRI acquisition is essential for clinical efficiency and improved patient experience.

Purpose of the Study:

• To introduce Self-Supervised Adversarial Diffusion for MRI Accelerated Reconstruction (SSAD-MRI), a novel deep learning approach.
• SSAD-MRI aims to significantly accelerate MRI data acquisition.
• The method is designed for accelerated reconstruction without relying on fully sampled reference datasets.

Main Methods:

• Utilized fastMRI multi-coil T2-weighted and single-coil T1 maps datasets for training and testing.
• Evaluated robustness using out-of-distribution (OOD) datasets (multi-coil T1c and T1-weighted).
• Data were retrospectively subsampled at acceleration rates R = 2x, 4x, and 8x; compared SSAD-MRI with ReconFormer and SS-MRI using NMSE, PSNR, and SSIM metrics.

Main Results:

• SSAD-MRI demonstrated superior preservation of fine structures and abnormalities compared to other methods at 8x acceleration.
• Achieved the lowest NMSE at 4x and 8x, and highest PSNR/SSIM across all rates for multi-coil data.
• Showed significant improvements (p ≪ 10^-5) in undersampled image quality on OOD datasets after reconstruction.

Conclusions:

• SSAD-MRI effectively reconstructs fully sampled images without using them during training.
• This self-supervised approach has the potential to reduce MRI imaging costs.
• SSAD-MRI enhances image quality, which is critical for accurate diagnosis and effective treatment planning.