Enhancing Interpretability in CT Reconstruction Using Tomographic Domain Transform with Self-supervision

Computed tomography (CT) reconstruction faces difficulties in dealing with artifacts caused by imperfect imaging processes. Deep learning-based CT reconstruction models have been proposed to address these challenges, but they often lack interpretability since they use end-to-end neural networks to directly transform signals from sinograms to CT sections. Additionally, supervised methods are commonly used to guide network training, yet obtaining supervision information can be challenging in biomedical imaging systems. To overcome these limitations, we propose a new domain transform CT reconstruction framework that includes self-supervision. Our approach interprets sinogram signals as tomographic information in the CT section domain, which is then used to formulate pixel intensities with a simple mean operation. A refinement network is utilized to improve the quality of the CT images, which are further processed using the Radon transform to achieve simulated sinograms without requiring additional supervision. Our experimental results demonstrate the effectiveness of the proposed framework in both anatomical structure reconstruction and artifact reduction.