CT-Mamba: A hybrid convolutional State Space Model for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Imaging

Background:

• Low-dose CT (LDCT) reduces radiation but introduces noise and artifacts.
• Existing deep learning denoising methods (CNNs, Transformers) have limitations in modeling long-range dependencies or computational complexity.
• Denoising can alter noise distribution, impacting diagnostic outcomes.

Purpose of the Study:

• To propose CT-Mamba, a hybrid convolutional State Space Model for effective LDCT image denoising.
• To address limitations of current denoising techniques by integrating local and global feature extraction.
• To ensure denoised images maintain noise characteristics similar to normal-dose CT (NDCT) images.

Main Methods:

• Developed CT-Mamba, a hybrid model combining CNNs for local features and Mamba for long-range dependencies.
• Implemented a spatially coherent Z-shaped scanning scheme for pixel continuity.
• Designed a Mamba-driven deep noise power spectrum (NPS) loss function for training.

Main Results:

• CT-Mamba demonstrated excellent performance in noise reduction and detail preservation in LDCT images.
• The model optimized noise texture distribution, achieving higher statistical similarity with NDCT images.
• Denoised images showed enhanced radiomics feature similarity to NDCT images.

Conclusions:

• CT-Mamba effectively denoises LDCT images, preserving details and improving noise texture.
• The hybrid approach balances local and global feature extraction for superior performance.
• CT-Mamba shows promise for Mamba framework application in LDCT denoising, enhancing diagnostic value.