FSformer: A combined frequency separation network and transformer for LDCT denoising

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) offers reduced radiation exposure but suffers from significant noise and artifacts.
• Image quality degradation in LDCT hinders accurate medical diagnosis.
• Existing denoising methods struggle to balance noise reduction with preservation of fine details.

Purpose of the Study:

• To develop an advanced deep learning model for superior low-dose computed tomography (LDCT) image denoising.
• To improve the clarity and diagnostic accuracy of LDCT images.
• To introduce a novel network architecture combining frequency separation and Transformer mechanisms.

Main Methods:

• Proposed a Frequency Separation network and Transformer (FSformer) model for LDCT denoising.
• Employed frequency separation blocks to decompose images into low- and high-frequency components.
• Utilized a Transformer stage for noise estimation and a reconstruction prediction block for image enhancement.
• Implemented a compound loss function incorporating frequency and Charbonnier loss.

Main Results:

• FSformer achieved state-of-the-art performance on AAPM Mayo, Piglet, and clinical datasets.
• Demonstrated optimal metrics on the Mayo dataset with PSNR of 33.77 dB and SSIM of 0.9254.
• Showcased effective noise/artifact suppression while preserving image texture and organization.
• Achieved a testing time of 1.825 seconds, indicating computational efficiency.

Conclusions:

• FSformer represents a significant advancement in LDCT image quality enhancement.
• The model effectively suppresses noise and artifacts, crucial for improved diagnostic accuracy.
• FSformer exhibits robustness and potential for widespread clinical application in medical imaging.