Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) signal acquisition involves capturing data in the frequency space, which is a corrupted Fourier transform of the desired image.
• Current deep learning methods for MRI reconstruction often process frequency and image space features separately, limiting their effectiveness.

Purpose of the Study:

• To introduce novel neural network layers that explicitly combine frequency and image feature representations for MRI reconstruction.
• To demonstrate the versatility of these joint learning schemes for various MRI reconstruction tasks.

Main Methods:

• Proposed neural network layers that integrate frequency and image feature representations.
• Developed joint learning schemes for artifact correction in frequency space and image manipulation.
• Applied joint convolutional learning to tasks including motion correction, denoising, and undersampled reconstruction.

Main Results:

• Joint models consistently produced high-quality output images across diverse tasks and datasets.
• Integration into unrolled optimization networks significantly improved the optimization landscape.
• Achieved an order of magnitude reduction in training time for undersampled MRI reconstruction.

Conclusions:

• Joint frequency and image feature representations are highly effective for MRI signal processing within deep learning networks.
• The proposed architecture offers a versatile building block for advanced MRI reconstruction.
• Publicly available code and models facilitate further research and application.