Dual Generative Adversarial Networks with Contextual-Dropout and Pairwise-Terms for Dual-Phase CT Scans Synthesis in Rectal Cancer

Multi-phase CT scanning has been proven to be a powerful imaging modality for colorectal imaging, commonly used for the diagnosis and evaluation of rectal cancer and plays a crucial role in preoperative localization, metastasis assessment, and postoperative follow-up of rectal cancer. However, multiphase CT images are acquired sequentially after contrast agent injection, resulting in significant local deformations between the images due to respiratory and cardiac motion. And the current common registration methods hinder the effective implementation of computer-aided diagnosis using fused multi-phase CT data. In this paper, we present a dual-phase CT scan registration method to establish anatomical correspondence between multi-phase CT images, thereby enhancing the potential for conducting relevant research on rectal cancer diagnosis through the fusion of multiphase CT data. Specifically, we assess the feasibility of utilizing contextual-dropout and paired-item combination training in dual generative adversarial networks to improve the authenticity of the synthesized results obtained from the dual-phase CT scan synthesis approach. We validated our method using a set of dual-phase CT scan data comprising 107 patients with rectal cancer. The experimental results demonstrated that the proposed approach effectively reduces the gap between synthesized and real dual-phase CT scans. This not only achieves accurate registration of dual-phase CT scan data in an alternative data synthesis manner but also enhances the potential of utilizing fused data from multi-phase CT for downstream computer-aided diagnostic tasks.