A new visual State Space Model for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but suffers from noise and artifacts impacting clinical utility.
• Convolutional Neural Networks (CNNs) and Transformers are common for LDCT denoising but have limitations in modeling capabilities or computational complexity.

Purpose of the Study:

• To develop a simple, efficient LDCT denoising model with linear computational complexity.
• The model aims to effectively capture both local spatial context and long-range dependencies.

Main Methods:

• Introduction of the Visual Mamba Encoder-Decoder Network (ViMEDnet), applying State Space Models to LDCT denoising.
• Proposal of the Mixed State Space Module (MSSM) combining depth-wise convolution, max-pooling, and 2D Selective Scan Module (2DSSM) for local and global feature extraction.
• Utilizing a weighted gradient-sensitive hybrid loss function to preserve image details during denoising.

Main Results:

• ViMEDnet demonstrated superior visual quality and quantitative performance compared to five state-of-the-art methods.
• The model effectively removed noise and artifacts while preserving fine structures and low-contrast edges.
• Quantitative metrics showed ViMEDnet achieved the lowest RMSE and highest PSNR, SSIM, and FSIM scores.

Conclusions:

• ViMEDnet offers excellent performance for LDCT denoising.
• It presents a viable alternative to existing CNN and Transformer-based denoising models.