Transformer-based arterial spin labeling perfusion MRI denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Arterial Spin Labeling (ASL) perfusion MRI quantifies cerebral blood flow (CBF) non-invasively.
• ASL MRI has a low signal-to-noise ratio (SNR), challenging image quality with limited data.
• Existing convolutional neural network (CNN) denoising methods can lose image texture and intensity variability.

Purpose of the Study:

• To develop an advanced denoising method for ASL CBF images.
• To address the limitations of current CNN-based denoising techniques.
• To improve the visualization and preservation of texture in ASL CBF images.

Main Methods:

• Proposed a Hybrid U-Net and Swin Transformer (HUST) model for ASL CBF denoising.
• Utilized U-Net as the backbone, incorporating Swin Transformers to replace CNN layers.
• Swin Transformers were employed for efficient feature extraction with reduced parameters via hierarchical structure and shifted window attention.

Main Results:

• HUST demonstrated substantial improvement in ASL CBF image visualization and texture preservation.
• The method was trained and tested on both 2D (277 subjects) and 3D (110 subjects) ASL CBF datasets.
• HUST achieved superior performance compared to three state-of-the-art methods, with higher PSNR and SSIM values for both 2D and 3D data.

Conclusions:

• HUST effectively denoises ASL CBF images, enhancing visualization and preserving essential image characteristics.
• The method allows for reduced data acquisition time without compromising CBF quantification quality.
• HUST represents a significant advancement in ASL perfusion MRI, offering superior image quality and efficiency.