DONet: Dual-Octave Network for Fast MR Image Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance (MR) image acquisition is time-consuming.
• Accelerating MR imaging is crucial for clinical applications.
• Parallel imaging techniques are commonly used to speed up acquisition by using undersampled data.

Purpose of the Study:

• To propose a novel deep learning framework, the Dual-Octave Network (DONet), for accelerated parallel MR image reconstruction.
• To leverage multiscale spatial-frequency features from both real and imaginary components of MR data.
• To improve the efficiency and accuracy of MR image reconstruction.

Main Methods:

• Developed a Dual-Octave Network (DONet) utilizing dual-octave convolutions (Dual-OctConvs).
• Each Dual-OctConv processes real and imaginary components, splitting them by spatial frequencies.
• Employed dense connections for feature reuse and aggregated contextual information across frequency groups.

Main Results:

• DONet demonstrated superior performance in accelerated parallel MR image reconstruction.
• The model effectively fused real and imaginary components across spatial frequencies.
• Experiments on clinical knee and fastMRI datasets validated the model's effectiveness under various acceleration factors and undersampling patterns.

Conclusions:

• DONet offers enhanced representational capacity by integrating multiscale spatial-frequency information.
• The network architecture facilitates efficient feature propagation and reuse.
• DONet significantly advances the field of accelerated MR image reconstruction, enabling faster and potentially more detailed imaging.