Transformer based deep learning denoising of single and multi-delay 3D Arterial Spin Labeling

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• 3D Arterial Spin Labeling (ASL) is crucial for non-invasive brain imaging.
• Denoising is essential to improve the signal-to-noise ratio (SNR) and accuracy of ASL data.
• Current deep learning methods, like CNNs, have limitations in effectively denoising complex ASL data.

Approach:

• Developed and evaluated Swin Transformer and CNN-based deep learning models for single-delay and multi-delay 3D ASL denoising.
• Trained and tested models on diverse datasets from multiple vendors to ensure generalizability.
• Assessed performance using similarity metrics, SNR, and accuracy in quantifying cerebral blood flow (CBF) and arterial transit time (ATT).

Key Points:

• Swin Transformer models outperformed CNNs for both single- and multi-delay 3D ASL denoising.
• Pseudo-3D Swin Transformer models with 3 slices offered an optimal balance between image quality and quantification accuracy.
• Spatiotemporal denoising models demonstrated superior performance for multi-delay ASL, reducing biases in CBF and ATT maps.

Conclusions:

• Swin Transformer deep learning models represent a significant advancement in denoising 3D ASL data.
• The proposed models enhance image quality and offer potential for reduced scan times.
• These improvements facilitate broader clinical adoption and application of 3D ASL.