Abstract
BACKGROUND: Low-dose computed tomography (LDCT) has become clinically essential for reducing radiation risks, but image noise remains a major challenge. Although deep learning methods have shown promise in denoising, they often fail to preserve fine structures while suppressing noise, particularly in orthopedic imaging where subtle bone textures are diagnostically critical. This study aimed to develop a wavelet-domain frequency-mixing transformer (FMT) network that simultaneously addresses noise suppression and structural preservation in orthopedic LDCT images, overcoming the limitations of current spatial-domain methods. METHODS: We developed a novel unfolding network integrating (I) multiscale wavelet decomposition for frequency band-specific processing; (II) FMT blocks for cross-band feature interaction; and (III) physics-based noise modeling for realistic denoising. The model was trained and validated with clinical orthopedic scans from medical centers, with quantitative evaluation according to peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). RESULTS: Our method achieved superior performance compared to six state-of-the-art approaches, with a mean PSNR of 42.3 dB (improvement of 3.7 dB over baseline) and an SSIM of 0.961 on the test data. Clinical evaluation by three radiologists confirmed the significantly better preservation of trabecular bone patterns (P<0.01). CONCLUSIONS: The proposed network establishes a new paradigm for LDCT denoising by explicitly modeling frequency-domain characteristics, demonstrating particular value for orthopedic applications requiring fine structural fidelity. This approach may enable further dose reduction in musculoskeletal imaging without compromising diagnostic quality.