HCformer: Hybrid CNN-Transformer for LDCT Image Denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose computed tomography (LDCT) reduces patient radiation exposure but introduces image noise, hindering diagnostic accuracy.
• Existing deep learning denoising methods, primarily Convolutional Neural Networks (CNNs), struggle with modeling complex structures due to their focus on local information.
• Transformer models offer global context but are computationally intensive for medical image processing.

Purpose of the Study:

• To develop an advanced image post-processing method for high-quality LDCT image reconstruction.
• To combine the strengths of CNNs and Transformers to overcome limitations in current denoising techniques.
• To improve the precision of clinical diagnosis from LDCT scans.

Main Methods:

• A hybrid CNN-Transformer (HCformer) codec network was proposed for LDCT image denoising.
• A neighborhood feature enhancement (NEF) module was designed to integrate local information into the Transformer.
• Shifting window multi-head self-attention (SW-MSA) and window multi-head self-attention (W-MSA) were alternately used to reduce computational complexity and enhance inter-layer information interaction.

Main Results:

• The HCformer model significantly improved image quality metrics: SSIM increased from 0.8017 to 0.8507, HuRMSE decreased from 34.1898 to 17.7213, and FSIM increased from 0.6885 to 0.7247.
• The algorithm successfully reduced noise while preserving crucial image details.
• Ablation and comparative experiments on the AAPM 2016 LDCT grand challenge dataset confirmed HCformer's superior performance over other methods.

Conclusions:

• The proposed HCformer effectively combines CNN and Transformer architectures for superior LDCT image denoising.
• HCformer demonstrates significant potential for enhancing medical image quality in low-dose CT applications.
• The method offers a promising solution for improving diagnostic accuracy in LDCT by providing high-quality, denoised images.