Abstract
X-ray acoustic computed tomography (XACT) suffers from low SNR and limited-view detection when only a limited number of detector channels are available. To address these limitations, we introduce a two-stage limited-view signal recovery framework that integrates frequency-aware denoising (FAD) diffusion with geometry-integrated masked autoencoders (GIMAE). FAD jointly models temporal and spectral noise characteristics to restore clean RF measurements, while GIMAE leverages detector-layout-guided masking to infer missing angular channels and reconstruct full-view RF sinograms from limited-view inputs. The recovered signals are subsequently used for XACT image reconstruction via time-reversal algorithm. Simulation and experimental evaluations using multiple X-ray irradiated patterns demonstrate substantial improvements in reconstruction fidelity, with the proposed method boosting SSIM by 31.1% in simulation and by 26.4% in experiments-closely matching full-view references and outperforming limited-view acquisitions. This framework provides an effective and practical solution for high-quality limited-view XACT imaging under realistic noise-dominated conditions.