RFDNet: Robust Frequency-Based Denoising Network for 3D Ultrasound Vascular Imaging Using a Row-Column Addressed...

Three-dimensional (3D) ultrasound vascular imaging (UVI) is essential for visualizing complex vascular structures. Row-column addressed (RCA) arrays, widely used for 3D UVI due to their hardware efficiency, suffer from point spread function (PSF) anisotropy, resulting in ramp-shaped noise that degrades image quality. Although existing denoising methods, including deep learning-based approaches, have shown promise, they are often limited by domain shift bias and the need for condition-specific data collection. Moreover, as full-volume 3D training is often impractical, many studies rely on 2D slice-wise training with 3D reconstruction, which can yield inter-slice intensity inconsistencies when slices are normalized independently. To overcome these limitations, we propose Robust Frequency-based Denoising Network (RFDNet), which integrates a Deep Frequency Filtering (DFF) module into a standard denoising model. The DFF module adaptively filters frequency components within the encoder, suppressing ramp-shaped noise while dynamically balancing spectral content to reduce sensitivity to domain shifts and inter-slice intensity inconsistencies. This adaptive filtering preserves vascular details and improves overall imaging consistency. Experiments on Doppler phantom, carotid artery, and abdominal datasets show that RFDNet significantly outperforms conventional methods in peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and root mean squared error (RMSE). Further validation through 2D frequency spectrum analysis confirmed that the DFF module dynamically adjusts frequency components to maintain spectral balance. In addition, spectral KL divergence analysis demonstrated its robustness against inter-slice intensity inconsistencies introduced by slice-wise normalization. This approach improves domain generalization, reduces noise artifacts, and enhances clinical applicability by improving imaging reliability. Future work will explore 3D training and architectural refinements for better computational efficiency.