Attention-guided frequency and locality learning GAN for limited-angle CT reconstruction

Computed tomography (CT) is a widely used technique with applications ranging from medical imaging to material analysis. One major challenge comes from the lack of scanning information at certain angles, resulting in distortion and artifacts in the reconstructed images. This problem known as Limited Angle Computed Tomography (LACT) reconstruction problem. To mitigate these challenges, we introduce a deep learning–based reconstruction framework that jointly incorporates frequency-aware and locality-aware mechanisms. The proposed model first restores corrupted spectral information by decomposing features into multiple frequency bands and applying learnable masks in the Fourier domain to selectively recover missing components caused by limited-angle sampling. Additionally, spatial consistency is enhanced through a locality-guided attention mechanism that leverages learned importance maps to weight query, key, and value computations, allowing the model to focus on structurally vulnerable regions requiring stronger correction. These components are integrated into a U-Net–based GAN architecture and optimized using a hybrid loss function combining adversarial, perceptual, frequency-domain, and sinogram-domain consistency terms. Experimental evaluation on the Mayo Clinic dataset demonstrates that our method outperforms existing reconstruction approaches, achieving superior PSNR and SSIM while producing visually cleaner and more anatomically consistent results.