Training deep learning based dynamic MR image reconstruction using open-source natural videos.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Dynamic MR imaging requires efficient reconstruction techniques.
• Data scarcity and sharing limitations hinder deep learning model development.
• Natural videos offer a potential alternative data source for training.

Purpose of the Study:

• To develop and evaluate a deep learning pipeline for dynamic MR image reconstruction using natural videos.
• To compare reconstruction performance using different DL architectures and sampling patterns.
• To assess the feasibility of training DL models with simulated MR data derived from natural videos.

Main Methods:

• Trained deep learning networks (VarNet, 3D UNet, FastDVDNet) with cardiac and natural video-derived synthetic MR data.
• Reconstructed real-time undersampled dynamic MR images using trained DL networks and compressed sensing.
• Evaluated performance using simulation metrics (MSE, PSNR, SSIM) and prospective assessments (image quality ranking, SNR, edge sharpness).

Main Results:

• Deep learning models trained with cardiac data outperformed those trained with natural videos in simulations.
• Both DL approaches significantly outperformed compressed sensing in simulations (p < 0.05).
• Prospective evaluations showed similar, superior rankings for DL reconstructions over compressed sensing, with no significant differences in SNR or edge sharpness.

Conclusions:

• A DL pipeline can learn dynamic MR reconstruction from natural videos, preserving high-quality, fast reconstructions.
• This approach mitigates limitations associated with data scarcity and sharing in MR imaging.
• The developed dataset, code, and networks are publicly available to advance research.