FMDNet: Spatial-frequency feature routing for low-dose CT denoising.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but introduces noise and artifacts.
• Existing deep learning denoisers may over-smooth images or create unstable textures.

Purpose of the Study:

• Develop an LDCT denoising network (FMDNet) with dedicated pathways for coarse structures and fine details.
• Improve noise suppression while preserving high-frequency anatomical information.

Main Methods:

• Propose FMDNet, a frequency-aware encoder-decoder network.
• Implement an explicit coarse/detail routing block using low-pass filtering and residual calculation.
• Refine components with depthwise operators and fuse using a gated channel attention mechanism.

Main Results:

• FMDNet achieved competitive quantitative fidelity on AAPM Mayo Clinic LDCT and LoDoPaB-CT datasets.
• Demonstrated improved PSNR, SSIM, and RMSE compared to baseline methods.
• Texture- and structure-oriented analyses confirmed enhanced detail preservation and local fidelity.

Conclusions:

• Explicitly separating coarse and detail components provides an effective bias for LDCT denoising.
• The FMDNet strategy enhances fidelity and preserves fine structures without adversarial synthesis.
• Further validation through clinical reader studies is recommended.