Scan-Specific Self-Supervised Bayesian Deep Non-Linear Inversion for Undersampled MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Magnetic resonance imaging (MRI) acquisition is limited by slow scan times due to data sampling constraints.
• Supervised deep learning (DL) accelerates MRI but requires extensive fully-sampled datasets.
• Existing unsupervised/self-supervised DL methods still necessitate large image databases.

Purpose of the Study:

• To introduce a novel scan-specific deep learning method for accelerated MRI reconstruction.
• To develop a method that eliminates the need for large datasets and calibration scans.
• To improve the performance of MRI reconstruction in calibrationless parallel imaging and compressed sensing.

Main Methods:

• Developed Scan-Specific Self-Supervised Bayesian Deep Non-Linear Inversion (DNLINV).
• Utilized a Deep Image Prior-type generative modeling approach.
• Employed approximate Bayesian inference for deep convolutional neural network regularization.

Main Results:

• Demonstrated improved performance across various anatomies, contrasts, and sampling patterns.
• Outperformed existing scan-specific calibrationless parallel imaging and compressed sensing methods.
• Successfully reconstructed MRI data using only sub-sampled data from a single scan.

Conclusions:

• DNLINV offers a powerful, self-supervised, scan-specific approach for accelerated MRI reconstruction.
• The method overcomes limitations of existing supervised and unsupervised DL techniques.
• DNLINV advances calibrationless parallel imaging and compressed sensing in MRI.