Computer Vision News - October 2022

49 Jiadong Zhang “Our framework uses dual-domain knowledge in two different modalities, PET and CT, and keeps dual-domain consistency in the synthesis process,” Jiadong explains. “To the best of our knowledge, ours is the first paper to exploit dual-domain information for two different modalities. Other works just use one modality, such as low-dose and standard-dose CT. We hope this work can make a significant contribution to cross-modality synthesis tasks.” The framework combines four neural networks with a forward projection and a filtered back projection, using bidirectional mapping with multiple closed cycles. These cycles can further serve as cycle-consistent constraints to keep the anatomical structures consistent in the synthesis process for better performance. “Synthesising CT image from PET image is challenging because some anatomical structures are not visible in PET image,” Jiadong tells us. “When using other medical image synthesis methods, such as CycleGAN, the anatomical structure in dual-domain is inconsistent, negatively impacting the results. Our work addresses this issue.” Jiadong points out that they have designed a general framework that could be used for many other applications. “So far, we have only explored our framework on PET-CT synthesis,” he says. “Next, we want to explore the performance