PIDA-GAN: physics-informed dual-stage attention GAN with data consistency for sparse-view CT reconstruction

Abstract Objective. Sparse-view CT (SVCT) reconstruction is an ill-posed inverse problem, where insufficient projection data leads to severe noise and streak artifacts. Although deep learning methods can mitigate these artifacts, they frequently oversmooth fine anatomical structures and lack explicit consistency with the measured data. This work aims to address these limitations by introducing a physics-informed dual-stage generative adversarial framework that enforces data consistency (DC) while preserving structural details and perceptual realism. Approach. We propose PIDA-GAN, a dual-stage attention GAN consisting of two cascaded U-Net–based generators enhanced with locality-guided self-attention and dynamic adaptive gating modules for context-aware feature refinement. The first stage produces an initial coarse reconstruction, which is subsequently refined through a physics-informed DC layer that performs iterative updates using forward and backprojection operators to reduce mismatch with the measured sinogram. The output is then enhanced by a second-stage generator trained jointly with the first stage using an edge-aware discriminator. The training process is guided by a hybrid loss function combining pixel-wise, structural similarity, perceptual, and adversarial losses. Main results. Experimental results on the Mayo Clinic dataset demonstrate that PIDA-GAN consistently outperforms state-of-the-art methods in terms of PSNR, SSIM, and RMSE across 60-, 30-, 16-, and 4-view settings, with substantial improvements observed under extreme sparsity. Qualitative evaluations further highlight its effectiveness in removing artifacts, reducing oversmoothing, and preserving structural fidelity. Moreover, experiments on the LIDC-IDRI dataset confirm strong cross-dataset generalization without fine-tuning. Significance. PIDA-GAN provides a robust physics-informed framework for ultra SVCT reconstruction with strong potential for clinical low-dose CT applications.