LUCMT: Learnable under-sampling and reconstructed network with cross multi-head attention transformer for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Compressed sensing MRI accelerates image acquisition but faces challenges in sampling strategy design and image reconstruction from under-sampled data.
• Existing methods often use static sampling patterns, limiting reconstruction performance.

Purpose of the Study:

• To develop an innovative approach, LUCMT, that simultaneously addresses sampling strategy design and image reconstruction for compressed sensing MRI.
• To improve the accuracy and efficiency of accelerated MRI.

Main Methods:

• LUCMT integrates a learnable under-sampling strategy with a Cross Multi-head Attention Transformer-based reconstruction network.
• An adaptive sampling scheme is learned via binarizing sampling patterns and backpropagation.
• The reconstruction network utilizes a CS-MRI depth unfolding approach with a Cross Multi-head Attention module.

Main Results:

• LUCMT demonstrated superior performance on T1 brain MR images from the FastMRI dataset compared to state-of-the-art methods.
• The method achieved high quantitative metrics (PSNR/SSIM) across various sampling rates (10-40%).
• LUCMT provided more accurate details and better visual quality in reconstructed images.

Conclusions:

• The proposed LUCMT method offers a promising solution for generating optimal under-sampling masks.
• LUCMT enables accurate and accelerated MRI reconstruction, advancing the field of medical imaging.