FMDNet: Spatial‐frequency feature routing for low‐dose CT denoising

BACKGROUND: Low-dose computed tomography (LDCT) is widely used to reduce radiation exposure, but the reduced photon budget amplifies quantum noise and can introduce structured artifacts that obscure subtle boundaries and textures. Many deep learning denoisers process features in a single stream, which may encourage either over-smoothing of weak anatomical edges or unstable texture synthesis. PURPOSE: To develop an LDCT denoising network that explicitly routes coarse structural content and fine details through dedicated pathways, aiming to suppress noise while preserving anatomically meaningful high-frequency information. METHODS: We propose FMDNet, a frequency-aware encoder-decoder equipped with an explicit coarse/detail routing block. A fixed low-pass operator produces a coarse component, while the corresponding detail component is formed as an explicit residual (high = x-low). Each component is refined with depthwise operators and fused by a learned channel gate with channel attention. We evaluate FMDNet on the AAPM Mayo Clinic LDCT dataset and the LoDoPaB-CT benchmark under a reproducible HU-windowed protocol and supplement PSNR, SSIM, and RMSE with texture- and structure-oriented analyses including HU line profiles and noise power spectrum (NPS), together with volume-level paired tests and bootstrap confidence intervals on LoDoPaB-CT. RESULTS: Across both benchmarks, FMDNet achieves competitive quantitative fidelity and shows favorable results on complementary structure-/texture-oriented analyses. On Mayo, it improves mean PSNR, SSIM, and RMSE relative to strong learned baselines. On LoDoPaB-CT, volume-level analysis across 28 validation volumes confirms consistent improvements over Uformer with paired tests and bootstrap confidence intervals. Additional HU-profile and NPS analyses provide complementary evidence of improved texture preservation and local structural fidelity relative to comparative methods. CONCLUSIONS: Explicitly separating coarse and detail residual components in feature space provides a practical inductive bias for LDCT denoising. When combined with gated fusion and multi-scale supervision, this strategy improves quantitative fidelity and preserves fine structures without relying on adversarial texture synthesis; clinical diagnostic impact should be validated by reader studies.