SIRE-Net: Structure-Integrated Refinement and Enhancement Network for Limited-Angle CT Reconstruction

Limited-Angle Computed Tomography (LACT) is an imaging technique suitable for plate-type batteries, enabling rapid projection data acquisition and image reconstruction within the range of angles where the battery is easily penetrable. However, the projection data acquired under limited-angle conditions, due to its inherent incompleteness, gives rise to considerable artifacts and distortions in the reconstructed images, which undermine both image quality and diagnostic accuracy. To address this challenge, this paper proposes the Structure-Integrated Refinement and Enhancement Network (SIRE-Net) for limited-angle CT reconstruction, specifically designed to recover lost image information and enable high-quality reconstruction from incomplete projection data. SIRE-Net is comprised of two primary sub-networks: the Structure-preserving and Artifact Removal Network (SRA-Net) and the Attention-based Contextual Enhancement Network (ACE-Net). Collectively, these two sub-networks constitute the generator of SIRE-Net. Specifically, SRA-Net ensures the preservation of structural integrity and the elimination of artifacts during the image reconstruction process, while ACE-Net enhances image details through the integration of an attention mechanism. The WGAN-GP discriminator, utilizing an adversarial training framework, further steers the generator toward the production of high-quality reconstructed images that more closely adhere to the statistical properties of authentic images. To evaluate the effectiveness of the proposed method, a real battery dataset was constructed, and comprehensive experiments were conducted using this dataset. The experimental results demonstrate that SIRE-Net is capable of generating high-quality CT images under limited-angle scanning conditions. In particular, SIRE-Net exhibits exceptional performance in mitigating edge artifacts, preserving structural integrity, and enhancing image details.