An Attention-Based Generative Adversarial Network for Limited-Angle CT Reconstruction

Reconstructing high-quality computed tomography (CT) images from limited-angle projections is a challenging and ill-posed problem, often resulting in severe artifacts and loss of structural details. Traditional analytical methods, such as Filtered Back Projection (FBP), struggle with incomplete data, while existing deep learning approaches face limitations in generalization and reliance on extensive paired datasets. To address these challenges, we propose a novel Generative Adversarial Network (GAN)-based framework comprising a U-Net-inspired generator enhanced with residual blocks and selfattention mechanisms, coupled with a PatchGAN discriminator. The generator effectively captures long-range dependencies and structural features critical for artifact removal and reconstruction accuracy, while the PatchGAN discriminator enforces local texture realism. Additionally, a perceptual loss derived from a pretrained VGG network preserves fine anatomical details and high-level semantic consistency. Extensive evaluations on clinical datasets demonstrate the superiority of our method over state-of-the-art techniques. Quantitative metrics, including PSNR, SSIM, and MSE, confirm significant improvements, and qualitative results showcase the effective suppression of artifacts and recovery of fine structural details.