Improving parallel translation computed laminography reconstruction quality via projection extrapolation

Parallel Translation Computed Laminography (PTCL) provides a new avenue for plate‑like object detection. However, it suffers from data loss, particularly in short‑travel scanning, which produces severe artifacts and degrades image quality. To address this limitation, we conduct an in‑depth analysis of PTCL’s frequency‑domain data distribution, clarify the inverse relationship between missing data and scan travel, and establish a theoretical foundation for subsequent work. Then, we propose a U‑shaped Swin‑Transformer Generative Adversarial Network (UST‑GAN), incorporating a Sinusoidal Neural Ordinary Differential Equation (SNODE) block and a novel Skip Horizontal Connection (SHC) block for projection data extrapolation in short-travel PTCL. SNODE models incremental change in a sinusoidal manner, ensuring smoothness and accuracy in the extrapolated projection region and mitigating artifacts typical of discrete extrapolation methods. SHC facilitates efficient feature transfer between the encoder and decoder, maintaining spatial consistency between extrapolated and original regions. Experimental results on simulated and real datasets demonstrate that the proposed UST‑GAN outperforms other frameworks in projection extrapolation, substantially reducing image artifacts and surpassing preprocessing baselines across all evaluation metrics.