PIPN: Red neuronal inspirada en la física para la obtención de fase en imágenes de contraste de fase de rayos X por...

Propagation-based X-ray phase-contrast imaging (PB-XPCI) can produce high-resolution images of soft tissue. However, this usually requires extracting the phase shift from intensity measurement at a single propagation distance through phase retrieval-an underdetermined nonlinear inverse problem. Conventional single-distance phase retrieval methods usually rely on multiple approximation conditions. Deep learning (DL)-based phase retrieval methods often rely on high-quality data for training or lengthy physics model iterative computations to optimize network parameters. In order to surmount the aforementioned limitations, this study proposes a physics-inspired phase retrieval network for propagation-based X-ray phase-contrast imaging (PIPN) and an acceleration strategy for the PIPN. It can achieve phase retrieval based solely on a single approximation condition and a physics imaging model, without the need for any training data. Experiments demonstrate that the PIPN can quickly reconstruct high-quality retrieved phase projections by using the acceleration strategy, and remain stable under different propagation distances.