High-performance rapid MR parameter mapping using model-based deep adversarial learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Image Reconstruction

Background:

• Accelerated Magnetic Resonance (MR) imaging is crucial for efficient parameter mapping.
• Traditional reconstruction methods may struggle with maintaining image quality at high acceleration rates.
• Deep learning offers potential for improving MR image reconstruction.

Purpose of the Study:

• To develop and evaluate a deep adversarial learning-based approach for rapid and efficient MR parameter mapping.
• To enhance image sharpness and texture restoration in MR parameter maps.
• To ensure data consistency in the reconstruction process.

Main Methods:

• A framework combining end-to-end convolutional neural network (CNN) mapping and adversarial learning was proposed.
• CNN performed direct image-to-parameter mapping from undersampled MR images.
• Adversarial learning improved image sharpness and texture; MR physical models ensured data consistency.

Main Results:

• The adversarial learning approach achieved accurate T2 mapping at an acceleration rate of R=8 for brain and knee datasets.
• The method yielded lower errors and higher similarity to reference maps compared to conventional techniques.
• Quantitative metrics showed improved normalized root mean square error, structural similarity index, and Tenengrad measures.

Conclusions:

• The proposed framework integrating CNN mapping, adversarial learning, and physical model-based data consistency is effective for MR parameter mapping.
• This approach shows promise for rapid and efficient reconstruction of quantitative MR parameters.
• The method demonstrated superior performance in maintaining image texture and sharpness.