Transformer With Double Enhancement for Low-Dose CT Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) usage is increasing due to rising health concerns, necessitating advanced image processing algorithms.
• Current CT algorithms aim to reduce radiation harm and image noise from dose reduction, but CNN-based methods struggle with broad region features.
• The transformer framework's large receptive field offers potential for improved CT image denoising.

Purpose of the Study:

• To develop an efficient and effective end-to-end deep learning network for low-dose CT (LDCT) image denoising.
• To leverage the transformer architecture for enhanced feature extraction and long-range dependency modeling in CT images.
• To improve the quality of LDCT images, reducing the need for higher radiation doses.

Main Methods:

• A novel end-to-end low-dose CT (LDCT) denoising network, DEformer, is proposed, utilizing a transformer framework.
• The network features a main branch with an overlapping-free window-based self-attention transformer block for denoising.
• Dual side branches incorporate a double enhancement module to enrich edge, texture, and context information, expanding the receptive field.

Main Results:

• Experiments on abdomen, head, and chest LDCT datasets demonstrated DEformer's superior denoising performance compared to existing algorithms.
• The proposed network effectively reduces noise while preserving essential image details, leading to higher quality CT images.
• The compound loss function (MSE, MSP, SL) aided in generating denoised images closer to norm-dose CT (NDCT) standards.

Conclusions:

• The transformer-based DEformer network offers a significant advancement in LDCT image denoising.
• DEformer enhances image quality by effectively processing broad regions and capturing long-range dependencies.
• This approach holds promise for reducing patient radiation exposure while maintaining diagnostic accuracy in CT imaging.