Deep J-Sense: Accelerated MRI Reconstruction via Unrolled Alternating Optimization.

Area of Science:

• Medical Imaging
• Machine Learning
• Signal Processing

Background:

• Accelerated MRI reconstruction combines compressed sensing and deep learning.
• Current methods often rely on estimated coil sensitivity profiles or calibration data.
• Performance degrades with poor estimators or differing scan parameters.

Purpose of the Study:

• Introduce Deep J-Sense, a robust deep learning approach for accelerated MRI reconstruction.
• Improve reconstruction performance and robustness against varying scan conditions.

Main Methods:

• Developed Deep J-Sense, a deep learning algorithm based on unrolled alternating minimization.
• The algorithm refines both the magnetization (image) kernel and coil sensitivity maps.
• Utilized a subset of the knee fastMRI dataset for experimental validation.

Main Results:

• Deep J-Sense demonstrated increased reconstruction performance.
• The method provided significant robustness to varying acceleration factors.
• Robustness was also observed across different calibration region sizes.

Conclusions:

• Deep J-Sense offers a robust deep learning solution for accelerated MRI.
• The simultaneous refinement of image kernels and coil maps enhances performance.
• This approach addresses limitations of existing methods relying on fixed or estimated parameters.