Projection completion based on score-based generative model

Purpose Scatter correction is a critical issue in cone-beam CT reconstruction, as effective scatter suppression can significantly enhance image quality. The beam stop array (BSA) is a well-established scatter correction method that plays a vital role in practical applications. However, while removing scattered rays, the BSA also obstructs primary rays, leading to missing projection data and consequently degrading reconstruction quality. This study aims to leverage generative models to specifically restore missing projection data, thereby improving reconstruction quality. Methods A score-based generative model was employed to train on scanned objects, learning the image distribution characteristics of the reconstructed subject. During multi-step sampling, domain transformation was integrated to convert generated image information into the projection domain. A weighted summation of known projections and generated projections was performed, followed by inverse domain transformation back to the image domain for subsequent sampling, ultimately yielding the final reconstructed image. Results This study proposes a novel approach for projection completion to address missing data caused by the BSA, with preliminary validation conducted on two independent datasets. Conclusion The proposed projection completion algorithm, incorporating a score-based generative model, demonstrates the capability to effectively restore missing projections in scenarios where the BSA induces significant data loss, thereby enhancing overall image quality.