MOST-DL

MOST-DL produces motion-robust T2 maps from MRI data by training neural networks on synthetic and in vivo data to mitigate motion artifacts and synthetic-to-real domain gaps.

Key Features:

• Synthetic Data Utilization: Generates synthetic MRI data via simulations that mimic real-world MRI conditions to augment training and mitigate scarcity of real samples.
• Calibrationless Parallel Reconstruction: Implements calibrationless parallel reconstruction using ultra-fast pulse sequences for rapid image acquisition without calibration.
• Intra-shot Motion Correction: Performs intra-shot motion correction to remove motion artifacts within single-shot acquisitions.
• Domain Gap Bridging: Incorporates textures from public databases and simulates imaging imperfections to reduce the synthetic-to-real domain gap during training.
• Motion-Robust T2 Mapping: Produces T2 maps resilient to unpredictable subject motion, reducing ghosting and motion artifacts in reconstructed images.
• Neural-network Training: Trains neural networks on synthetic datasets to enable generalization to real MRI data.

Scientific Applications:

• Brain MRI T2 mapping: Validated on simulation and in vivo human brain MRI data to improve image quality by minimizing head motion artifacts.
• Clinical imaging of motion-prone patients: Applicable to clinical imaging workflows to reduce ghosting and motion artifacts in patients who move during scans.
• General medical imaging: Applicable to other medical imaging scenarios where motion artifacts impede quantitative mapping.

Methodology:

Neural networks are initially trained on synthetic MRI data (including simulated imperfections and textures from public databases) and evaluated on in vivo human brain scans.