Deep Prior Based Limited-Angle Tomography

In the process of reconstructing images from data acquired within a limited angular range, we encounter what is termed limited-angle tomography. The deficiency of complete data in this context results in artifacts, commonly appearing as streaks or missing structures, which can significantly compromise the quality of the reconstructed slice. This degradation gives rise to issues such as boundary distortion, blurred edges, and intensity bias, potentially leading to misinterpretation of the images. Hence, addressing artifacts in limited-angle tomography is crucial for clinical applications. Although deep learning-based reconstruction has shown impressive results in recent times, concerns about its robustness persist. To bolster the robustness of our proposed technique, we integrate prior information from a modified U-net with preprocessed input into the Relative Variation - Simultaneous Algebraic Reconstruction Technique (RV-SART) to provide insights into unmeasured data. Subsequently, the method extracts structure from the initially reconstructed slice through structure-texture decomposition. This process facilitates the reconstruction of high-quality CT images while suppressing pattern-like artifacts. Extensive experiments demonstrate that our approach surpasses both traditional and state-of-the-art learning techniques in terms of reconstruction quality and preservation of fine structures in noisy limited-angle reconstruction problems. Our technique provides improvements over the recent LRIP-net for a 90-degree scanning range in quantitative metrics such as PSNR by 17.48%, RMSE by 46.36%, and SSIM by 6.18%.