Phase Retrieval Methods For In-Line X-Ray Phase Contrast Imaging Based On Deep Learning

X-ray Phase Contrast Imaging(XPCI)has the advantages of obtaining high contrast images at low radiation dose,and has shown great potential in clinical diagnosis.Among the currently developed XPCI method the imaging devices of in-line XPCI is simple and does not require additional optics except for the X-ray source and detector,making it been one of the XPCI methods closest to clinical applications.In practical in-line XPCI,the non-ideal factors such as the finite size of focal spot of X-ray source,finite pixel size of detector,and the system noise all affect the phase contrast.However,the current commonly used conventional phase retrieval methods,such as the Phase Attenuation Duality(PAD)method,usually build ideal and simplified model based on certain assumptions,therefore its practical application range are limited.In order to overcome the limitations of the conventional phase retrieval methods and improve the signal-to-noise ratio and spatial resolution of the retrieved phase projections,this thesis fully considers the non-ideal factors in the imaging process and establishes an accurate phase contrast imaging model,then the nonlinear mapping between phase contrast projection and the phase projection is built by the Generative Adversarial Networks(GAN)and concatenation networks based on the deep learning,which lays the foundation for more accurate phase tomography.The generative adversarial networks relies on the process of game between generator and discriminator for training.This thesis builds a network model based on the principle of GAN to phase retrieval of in-line XPCI.Solid and hollow sphere phantoms were designed to simulate human soft tissue and polyethylene bubble with weak absorption,respectively,under non-ideal imaging conditions,,and the phase retrieval results are quantitatively evaluated using structural similarity,peak signal-to-noise ratio and full width at half maximum.The results of hollow sphere simulation show that the spatial resolution and peak signal-to-noise ratio of the phase projection obtained by the phase retrieval method based on GAN are improved by an average of 42% and 33%,respectively,compared with the conventional PAD phase retrieval method.In order to reduce the training time and use a small amount of data to achieve the simultaneous increase of anti-noise capability and spatial resolution,a plug-and-play concatenation network is proposed.The network is composed of the phase contrast effect recovery module and the phase retrieval module,which reduces the influence of non-ideal factors in in-line XPCI on the phase retrieval.The network was validated by the phantom of lung histopathology and hollow sphere and compared to the conventional PAD method.In the case of small amount of training data,the results of hollow sphere phantom show that the spatial resolution and peak signal-to-noise ratio of phase projection retrieval by the network are improved by an average of 67% and 52% respectively.In terms of practical data testing,the phase contrast projections of polyethylene bubble were obtained by the in-line XPCI system in the laboratory,and the two phase retrieval networks proposed in this thesis are further verified.The results show that,compared with the conventional PAD phase retrieval method,the above two convolutional neural network methods have stronger ability to suppress noise,and the phase projection edge obtained by the deep concatenation networks is clearer.