SR-Net: A sequence offset fusion net and refine net for undersampled multislice MR image reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning for fast magnetic resonance imaging (MRI) reconstruction is popular.
• Reconstruction quality degrades with high acceleration factors in undersampled MRI.
• Exploring inter-slice data redundancy can improve undersampled MRI reconstruction.

Purpose of the Study:

• To enhance the reconstruction quality of undersampled MR images.
• To leverage spatial correlations among slices for improved MRI reconstruction.
• To address challenges posed by large acceleration factors in MRI.

Main Methods:

• Developed a novel deep learning framework with two subnets: Sequence Offset Fusion Net (S-Net) for inter-slice correlations and Refine Net (R-Net) for intra-slice correlations.
• Utilized deformable convolution for neighbor slice feature extraction in S-Net.
• Incorporated a data consistency operation and iterative application of S-Net and R-Net for dealiasing in the image domain.

Main Results:

• The proposed method demonstrated superior reconstruction results compared to state-of-the-art techniques on public MRI datasets.
• Achieved significant improvements in dealiasing and restoration of tissue structures.
• Demonstrated real-time processing capability with reconstruction speeds exceeding 14 slices per second for 256x256 images.

Conclusions:

• Spatial correlation among MRI slices serves as valuable prior information.
• The proposed method significantly enhances the reconstruction quality of undersampled MR images.
• The approach offers a viable solution for accelerating MRI acquisition while maintaining image fidelity.