Self-Attention Convolutional Neural Network for Improved MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) is crucial but suffers from long acquisition times.
• Accelerated MRI acquisition necessitates robust image reconstruction techniques.
• Deep learning methods show promise but face challenges with local receptive fields.

Purpose of the Study:

• To develop a deep learning framework for accelerated MRI reconstruction with enhanced image fidelity.
• To integrate self-attention mechanisms for improved signal synthesis and artifact compensation.
• To validate the framework on cartilage MRI data acquired with ultrashort echo time sequences.

Main Methods:

• Proposed a novel deep learning framework, SAT-Net, integrating self-attention modules into a hierarchical deep residual convolutional neural network.
• Employed dense shortcut connections and enforced data consistency for improved reconstruction.
• Utilized a volumetric network architecture applied to retrospectively undersampled cartilage MRI data.

Main Results:

• The SAT-Net demonstrated improved image fidelity for accelerated MRI reconstruction.
• Self-attention mechanism effectively captured long-range dependencies, enhancing signal synthesis.
• The framework achieved superior outcomes on cartilage MRI compared to conventional methods.

Conclusions:

• The proposed SAT-Net framework offers a promising solution for accelerating MRI acquisition while maintaining high image quality.
• Integration of self-attention mechanisms addresses limitations of purely convolutional approaches in MRI reconstruction.
• The generic framework is adaptable for diverse accelerated MRI applications.