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

BACKGROUND: Sparse-view CT reconstruction is an ill-posed inverse problem, where insufficient projection data leads to degraded image quality with increased noise and artifacts. Although deep learning methods can mitigate these artifacts, they frequently oversmooth fine structures and lack explicit consistency with the measured sinogram data. PURPOSE: We introduce PIDA-GAN, a physics-informed dual-stage attention generative adversarial network that integrates data-consistency constraints with adaptive feature learning to achieve accurate and perceptually realistic sparse-view CT reconstruction. METHODS: PIDA-GAN consists of two cascaded U-Net-based generators enhanced with locality-guided self-attention (LGSA) and dynamic adaptive gating (DAG) to enable context-aware feature refinement. The proposed architecture first produces an initial coarse reconstruction and then refines it in the second refinement stage. Furthermore, a physics-informed data consistency (DC) layer is added between the stages, performing iterative updates using TorchRadon forward and backprojections to reduce the mismatch with the measured sinogram. The second stage is trained jointly with the first stage using an edge-aware discriminator to enhance image sharpness and textures. The training process uses a hybrid loss function that consists of pixel-wise, structural similarity (SSIM), perceptual, and adversarial losses. RESULTS: Experimental results on the Mayo Clinic dataset show that PIDA-GAN outperforms state-of-the-art methods for ultra sparse-view CT reconstruction using 60-, 30-, 16-, and 4-view settings. The proposed method achieves the best performance in terms of PSNR, SSIM, and RMSE, with significant improvements observed under extreme sparsity. Qualitative evaluations demonstrate the effectiveness of PIDA-GAN in artifact removal, suppression of oversmoothing, and preservation of structural fidelity. CONCLUSION: PIDA-GAN provides a robust physics-informed framework for ultra sparse-view CT reconstruction, delivering accurate and artifact-suppressed images with strong potential for clinical low-dose CT applications.