Contrast-enhanced CT Image Synthesis of Thyroid Based on Transfomer and Texture Branching
Ning Xiao, Zhenyu Li, Shaobo Chen, Liangtian Zhao, Yuer Yang, Hao Xie, Yang Liu, Yujuan Quan, Junwei Duan · 2022
Thyroid cancer is one of the most common cancers in young women, but due to the noise and artifacts of ultrasound images, there is still a certain misdiagnosis rate in clinical practice, and it is often combined with plain and Contrast-enhanced CT for further diagnosis. Compared with plain CT, Contrast-enhanced CT has better contrast with reflecting the erosion of organ margins, which is an important symptom for diagnosing thyroid cancer. Contrast-enhanced CT, however, relies on the patient being injected with a contrast agent and exposed to ionizing radiation. Our work proposes an improved Unet architecture. To generate enhanced CT images with clear texture and higher quality, we use the convolutional Transformer module to learn the global information of high-dimensional features, and then fuse the texture feature module to extract the local texture information of plain CT and the edge information extracted by superpixels as a priori knowledge to restore texture details. Experimental results show that our framework outperforms state-of-the-art generative networks and can generate higher-quality Contrast-enhanced CT.