Unet-like transformer with variable shifted windows for low dose CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose computed tomography (LDCT) is essential for minimizing radiation exposure in medical imaging.
• LDCT images often suffer from noise and artifacts, which can reduce diagnostic accuracy.
• Standard Transformer models show promise for LDCT denoising but are computationally expensive for high-resolution images.

Purpose of the Study:

• To develop a computationally efficient Transformer-based model for low-dose computed tomography (LDCT) image restoration.
• To address the limitations of standard Transformers in handling high-resolution medical images.

Main Methods:

• A novel pure Transformer architecture integrated into a hierarchical U-Net framework was proposed.
• An agent attention mechanism was developed to approximate global self-attention with linear complexity.
• The model utilizes a variable shifted-window design to capture both local details and global context efficiently.

Main Results:

• The proposed method achieved state-of-the-art performance on a public LDCT dataset.
• Quantitative metrics and qualitative visual comparisons demonstrated superior denoising capabilities.
• The model effectively reduced noise and artifacts while preserving important image details.

Conclusions:

• The developed Transformer architecture offers an efficient solution for LDCT image denoising.
• The agent attention mechanism successfully balances computational efficiency and performance.
• This approach has the potential to improve diagnostic accuracy in low-dose CT imaging.