An unsupervised physics-informed neural deconvolution framework for ultrasound imaging

Abstract

Objective. Point spread function (PSF) and clutter noise are primary causes of ultrasound quality degradation. Physical deconvolution can reduce PSF effects but is sensitive to high-frequency noise, whereas deep learning often requires large labeled datasets and may introduce non-physical artifacts. Therefore, this study aims to develop an unsupervised and physically interpretable framework for robust and effective ultrasound image reconstruction. Approach. We propose an unsupervised physics-informed deconvolution network framework (UPIDN) that bridges the strengths of physical modeling and deep learning. Specifically, the degradation model is embedded into the network, and physical deconvolution provides stable low-frequency inversion guidance. Meanwhile, the structural prior capability of the deep image prior is exploited to compensate for the ill-posedness of high-frequency detail recovery. Additionally, we perform a cepstral domain transformation on radio-frequency data to decouple the embedded PSF information, providing the network with a physically consistent and data-adaptive initialization strategy. Meanwhile, considering the ultrasound noise characteristics, a dual-wavelet noise preprocessing scheme is designed to guide the prediction to focus more on the correct generation direction. Main results. Experiments demonstrate that UPIDN outperforms other state-of-the-art methods, achieving superior contrast and resolution while clearly measuring vessel–muscle interface that is difficult to discern in input delay-and-sum results. Significance. UPIDN provides a reliable approach for high-quality ultrasound reconstruction without paired training data and offers a potential solution for ultrasound imaging scenarios where data availability and physical fidelity are critical.