PSDP: Pseudo-supervised dual-processing for low-dose cone-beam computed tomography reconstruction

Low-dose cone-beam computed tomography (CBCT) is reconstructed from hundreds of 2D X-ray projections of low intensity to reduce ionizing radiation to patients, but its imaging quality suffers from noise degradation. A mainstream solution is the supervised dual-processing method to successively improve the quality of projections and reconstructed CT images based on large amounts of paired normal-dose and low-dose projections, which are often unavailable in clinical practice. Therefore, we present a Pseudo-Supervised Dual-Processing (PSDP) strategy which includes an unpaired-learning Degrading Projection Network (DeProjNet) and a dual-processing network. Specifically, our DeProjNet is implemented with a generative model, and it builds the projection pairs by degrading the normal-dose projections into their pseudo low-dose counterparts. In particular, we develop a two-path adversarial learning into DeProjNet to effectively learn the noise distributions from the real unpaired low-dose projections. To overcome the common challenge of the undesired content deformation while using a generative model, we develop a physics-based content-fidelity loss to constrain the image content of the generated pseudo low-dose projections. A well-performed dual-processing network is integrated with our DeProjNet to accurately reconstruct low-dose CT images. Experimental results demonstrate that our PSDP obtains the similar root mean square error (RMSE) as the existing paired-learning method, and remarkably decreases RMSE by approximately 16% compared to the existing unpaired-learning method. In conclusion, our PSDP is a promising method and may have wide applications for reconstruction of low-dose CBCT.