Coded Aperture Computed Tomography Via Generative Adversarial U-net

Generative adversarial U-net for coded aperture computed tomography (CT) is proposed in this paper to alleviate the tradeoff between the non-continuous sparse projections and the ill-posedness iterative reconstruction problem. A non-continuous sparse projection model is presented based on generative adversarial U-net and the corresponding joint penalty function is formulated. Simulations using real datasets show that CT images with 256×256 pixels can be reconstructed with peak signal-to-noise ration more than 30 dB at only 5% transmittance. Furthermore, the computational time in the reconstructions is reduced by two orders of magnitude when compared with the state-of-the-art iterative algorithms in coded aperture computed tomography.