Time-Dependent Deep Image Prior for Dynamic MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Dynamic MRI is crucial for imaging moving organs, but rapid acquisition is challenging.
• Existing methods often require extensive training data or specific preprocessing steps.

Purpose of the Study:

• To introduce a novel unsupervised deep learning algorithm for dynamic MRI reconstruction.
• To enable high-resolution reconstruction of dynamic MRI sequences without prior training or additional data.

Main Methods:

• A generalized deep-image-prior approach optimizing a reconstruction network's weights.
• Utilizing a fixed low-dimensional manifold for temporal variations and a CNN for image generation.
• Ensuring consistency with k-space measurements without heartbeat marking or spoke reordering.

Main Results:

• Outperformed state-of-the-art methods quantitatively and qualitatively on cardiac datasets.
• Successfully reconstructed continuous variations in dynamic MRI sequences.
• Achieved high spatial resolution in dynamic MRI reconstruction.

Conclusions:

• This unsupervised deep learning method offers a significant advancement in dynamic MRI reconstruction.
• The approach eliminates the need for prior training or specific cardiac data preprocessing.
• It enables high-fidelity imaging of dynamic physiological processes.