A multi-scale residual network for accelerated radial MR parameter mapping.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance (MR) parameter mapping is crucial for quantitative imaging.
• Accelerated data acquisition in MR is needed to reduce scan times and improve patient comfort.
• Conventional reconstruction methods can be time-consuming and may struggle with highly undersampled data.

Purpose of the Study:

• To develop and evaluate a deep learning framework for accelerated MR parameter mapping.
• To enable high-quality image reconstruction from radially undersampled MR data.
• To significantly reduce MR image reconstruction times for potential clinical applications.

Main Methods:

• A deep learning framework combining accelerated radial data acquisition with a multi-scale residual network (MS-ResNet) for image reconstruction.
• Supervised learning strategy using multi-contrast image patches with and without radial undersampling artifacts.
• Subspace filtering for denoising input patches and individual network training per anatomy and relaxation parameter.

Main Results:

• MS-ResNet outperforms conventional model-based compressed sensing methods for T2 mapping, even with undersampled training data.
• Achieved comparable contrast-weighted images and parameter maps to iterative methods for T1 and T2 mapping.
• Demonstrated a two-orders-of-magnitude reduction in reconstruction times for in vivo brain and abdomen T1 mapping, and brain T2 mapping.

Conclusions:

• The proposed MS-ResNet framework enables high-quality MR parameter mapping from highly accelerated radial acquisitions.
• The significant reduction in reconstruction time makes this approach suitable for routine clinical use.
• This deep learning method offers a promising solution for efficient and accurate quantitative MR imaging.