A cross-domain complex convolution neural network for undersampled magnetic resonance image reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) reconstruction from undersampled k-space data is crucial for faster scans.
• Existing deep learning methods face challenges in feature extraction across image and k-space domains.
• Current cross-domain approaches struggle with simultaneous feature extraction and fusion.

Purpose of the Study:

• To introduce TEID-Net, a novel cross-domain complex convolution neural network for accurate MR image reconstruction.
• To address limitations in feature extraction and artifact reduction in existing reconstruction methods.
• To enhance the quality of MR images reconstructed from undersampled k-space data.

Main Methods:

• Proposed a novel deep-learning-based 2-D single-coil complex-valued MR reconstruction network, TEID-Net.
• Integrated three modules: TE-Net (image-domain, Texture Enhancement Module), ID-Net (intermediate-domain, aliasing artifact reduction), and the cascaded TEID-Net.
• Utilized fastMRI and Calgary-Campinas datasets for extensive experimental validation.

Main Results:

• TEID-Net effectively mitigates undersampling-induced artifacts, producing high-quality MR image reconstructions.
• The proposed network outperforms several state-of-the-art methods in image reconstruction quality.
• TEID-Net demonstrates superior performance in restoring tissue structures and intricate texture details, with fewer parameters.
• The method is particularly effective for regular Cartesian undersampling scenarios.

Conclusions:

• TEID-Net represents a significant advancement in deep learning-based MR image reconstruction.
• The cross-domain approach effectively integrates features from both image and k-space domains.
• TEID-Net offers a promising solution for high-fidelity MR image reconstruction, especially in undersampled settings.