Spatial orthogonal attention generative adversarial network for MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Self-attention modules excel at capturing long-range dependencies in vision tasks.
• Existing self-attention methods for MRI reconstruction are often computationally expensive and memory-intensive.
• There is a need for efficient attention mechanisms tailored for MRI reconstruction.

Purpose of the Study:

• To design a lightweight spatial orthogonal attention module (SOAM) for capturing long-range dependencies.
• To develop a novel spatial orthogonal attention generative adversarial network (SOGAN) for accurate MRI reconstruction.
• To address the high computational complexity and memory usage of current attention modules in MRI.

Main Methods:

• Developed a lightweight SOAM generating small attention maps for vertical and horizontal long-range context aggregation.
• Integrated SOAMs into concatenated convolutional autoencoders to form the SOGAN generator.
• Evaluated SOGAN against state-of-the-art deep learning-based compressed sensing MRI (CS-MRI) methods.

Main Results:

• SOAMs effectively enhance reconstructed MR image quality by capturing long-range dependencies.
• SOGAN achieves more accurate MRI reconstruction compared to existing deep learning CS-MRI methods.
• SOGAN demonstrates superior performance with a reduced number of model parameters.

Conclusions:

• The proposed SOAM is an effective and lightweight module for capturing long-range dependencies, significantly improving MRI reconstruction quality.
• SOGAN, incorporating SOAMs, surpasses state-of-the-art deep learning-based CS-MRI methods in performance.
• The study presents a promising direction for efficient and accurate deep learning-based MRI reconstruction.