MANTIS: Model-Augmented Neural neTwork with Incoherent k-space Sampling for efficient MR parameter mapping.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Magnetic Resonance Imaging

Background:

• Quantitative MR parameter mapping is crucial for accurate diagnosis.
• Current methods face challenges with scan time and image quality, especially at accelerated acquisition rates.

Purpose of the Study:

• To develop and evaluate MANTIS (Model-Augmented Neural neTwork with Incoherent k-space Sampling), a novel deep learning approach for efficient MR parameter mapping.
• To assess MANTIS's performance in T2 mapping of the knee at various acceleration factors.

Main Methods:

• MANTIS integrates end-to-end convolutional neural network (CNN) mapping with incoherent k-space undersampling and a physical model.
• A CNN directly reconstructs MR parameter maps from undersampled images.
• A signal model pathway ensures synthesized k-space consistency with acquired data.

Main Results:

• MANTIS produced high-quality T2 maps at R=5 and R=8 acceleration rates.
• It demonstrated lower errors and higher similarity to reference scans compared to conventional and other CNN methods.
• Normalized RMSE was 6.1% (R=5) and 7.1% (R=8); SSIM was 86.2% (R=5) and 82.1% (R=8).

Conclusions:

• The MANTIS framework offers an efficient and robust method for quantitative MR parameter estimation.
• Its combination of CNN mapping, signal model augmentation, and undersampled k-space acquisition shows significant promise.
• MANTIS advances the field of accelerated quantitative MRI.