LAUSR

Dual-energy computed tomography (DECT) can simultaneously provide the anatomical structure and material-specific information of the scanned object, having many applications in industry and medicine. Different from conventional CT, DECT acquires two attenuation measurements of the same object at two different X-ray spectra, resulting in apparent redundant information. This article exploits this kind of redundancy to develop the self-prior information enhanced deep iterative reconstruction (SPIE-DIR) algorithm for limited-angle DECT. Unlike the routine practice in model-based deep learning (DL) algorithms, the SPIE-DIR method simultaneously performs constraints in the projection, residual, and image domains, corresponding to three modules: projection inpainting, residual correction, and image refinement. During this stage, the prior image and prior projection derived from two complementary limited-angle scans are used to improve the algorithm performance. Besides, to avoid the blurring effect caused by minimizing the Euclidean distance, the Wasserstein generative adversarial network with gradient penalty is adopted to enhance the visual perception of the generated results. Experiments on the simulated data and real rat data have demonstrated that the proposed SPIE-DIR algorithm has the potential to obtain high-quality DECT images from two limited-angle scans. Furthermore, visual and quantitative assessments have shown the promising performance of SPIE-DIR in artifact removal, structural fidelity, CT number preservation, and visual perception enhancement.