Unsupervised reconstruction of accelerated cardiac cine MRI using neural fields

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Cardiac cine MRI is essential for assessing heart function but slow acquisition necessitates reconstruction methods for accelerated, undersampled scans.
• Existing regularization and supervised deep learning methods accelerate MRI but require large datasets, potentially introducing bias.
• Unsupervised learning offers an alternative to overcome data limitations in cardiac MRI reconstruction.

Purpose of the Study:

• To introduce NF-cMRI, an unsupervised deep learning method for reconstructing cardiac cine MRI.
• To evaluate NF-cMRI's performance on in-vivo undersampled data at high acceleration factors.
• To demonstrate the potential of implicit neural fields for accelerated cardiac MRI.

Main Methods:

• Developed NF-cMRI, an unsupervised approach utilizing implicit neural field representations.
• Applied the method to in-vivo golden-angle radial multi-coil acquisitions.
• Tested performance at undersampling factors of 13x, 17x, and 26x.

Main Results:

• NF-cMRI achieved excellent sharpness and artifact robustness.
• The method provided spatial-temporal depiction comparable or superior to existing techniques.
• Demonstrated effectiveness in reconstructing highly undersampled cardiac cine MRI.

Conclusions:

• NF-cMRI shows significant potential for reconstructing cardiac cine MRI from highly undersampled data.
• The unsupervised, implicit neural field approach addresses limitations of data-dependent methods.
• This technique advances accelerated cardiac imaging possibilities.