Planning CT Guided Limited-Angle CBCT to CT Synthesis via Content-Style Decoupled Learning

Abstract

Cone-beam computed tomography (CBCT) images are usually applied to clinical tasks such as image-guided radiation therapy due to the capability of providing accurate anatomical structures of patients. CBCT data obtained by full-angle scan takes a long scanning time and has a relatively high radiation dose, which may increase the health risks and discomfort of patients. Limited-angle CBCT (LA-CBCT) can effectively decrease scanning time and radiation dose by reducing the scanning angle range. On the other hand, it suffers from serious wedge artifacts, loss of image details, and low Hounsfield unit (HU) accuracy. Hence it is worthwhile to investigate the generation of high-quality CT-like images from LA-CBCT. However, due to neglect of the recovering of missing anatomical content caused by the limited-angle scan, traditional DL-based methods fail to generate high-quality synthetic CT from LA-CBCT directly. To solve this problem, we make full use of the bidirectional mapping between CBCT and CT domain and decouple LA-CBCT to CT synthesis into image style (context, HU) learning stage and image content (anatomical structure) learning stage. To accurately correct the image texture and intensity, an edge-enhanced generative adversarial network (EEGAN) is proposed to learn the bidirectional mapping relationship between CBCT and CT images. To recover the missing content caused by the limited-angle scan, a prior-guided content supplement network (PGCS-Net) is proposed to eliminate the limited-angle artifacts and supplement the missing anatomic structure. Results on real clinical chest and head datasets indicate that synthetic CT generated with our method can effectively improve image quality and registration quality of LA-CBCT, and has great potential in some image-guided radiotherapy tasks such as patient setup error obtaining.