Reducing Switching Noise in Radiofrequency Ultrasound Signals Using Deep Neural Networks

Ultrasound (US) systems suffer from digital switching noise (SN) originating from power supplies. Typical switching frequencies are in the MHz range and often fall within the transducer's bandwidth, thus deteriorating image quality. Conventional notch filtering can attenuate SN but at the cost of losing valuable signal content. In this work, we present an experimental approach to acquire US radiofrequency (RF) signals affected by SN and their approximated noise-free counterparts, using the ULA-OP 256 research scanner with a mechanically translated linear probe. By realigning consecutive acquisitions, we exploited the coherence of SN across elements to suppress it through signal averaging, yielding a dataset of 35136 noisy ad noise-free signal pairs. This dataset was used to train a 1D convolutional encoder-decoder neural network to map the spectral magnitude of noisy RF signals to their denoised version. Quantitative evaluation demonstrated an up to around 4.0 dB reduction of SN components within the -20 dB signal bandwidth, as well as an improved contrast of in vivo B-mode images obtained from denoised signals. The proposed approach effectively reduces SN artifacts while preserving useful information and provides a foundation for further extensions to 2D network architectures and additional imaging scenarios.