[Physical model-based cascaded generative adversarial networks for accelerating quantitative multi-parametric...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Context:

• Quantitative multi-echo multi-parametric MRI is crucial for detailed tissue characterization.
• Accelerating MRI acquisition while maintaining image quality is a significant challenge.
• Generative Adversarial Networks (GANs) show promise in image reconstruction tasks.

Purpose:

• To investigate the feasibility of physical model-based cascaded GANs for accelerated quantitative multi-echo multi-parametric MRI.
• To develop a deep learning reconstruction method utilizing raw multi-echo multi-coil k-space data.
• To enhance image feature information and improve the quality of reconstructed MR images.

Summary:

• A novel physical model-based cascaded GAN was proposed, integrating multi-domain information and optimizing network structures for k-space and image generation.
• The method reconstructs high-quality images from raw multi-echo multi-coil k-space data, employing a system matrix for improved generalization.
• Quantitative evaluation on an 80-case test set demonstrated superior performance in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Normalized Root Mean Square Error (NRMSE) compared to existing methods.

Impact:

• The proposed GAN model significantly improves overall image quality and quantitative accuracy in multi-parametric MRI.
• Reconstructed images exhibit clearer visualization of anatomical structures like gray matter, white matter, and cerebrospinal fluid.
• This physically driven deep learning approach offers a substantial advancement in accelerating MRI acquisition without compromising diagnostic information.