Contrast Agent Removal for Brain CT Angiography Using Switchable CycleGAN with AdaIN and Histogram Equalization

Computed tomography angiography uses an injection of contrast agent into the blood vessels, and helps to diagnose diseases that occur in the vessels and soft tissues. This requires contrast-enhanced CT (CECT) and non-enhanced CT (NECT) images, and the rigid registration and bone subtraction techniques are applied for better vessel visualization. However, the visualization process needs additional radiation exposure for obtaining two scans for the CECT and NECT. Also, it has limitations in that some vessels can be partially deleted due to insufficient registration. In this work, we propose a method to synthesize NECT from CECT images, using cycle-consistent generative adversarial network (cycleGAN) with adaptive instance normalization. Especially, a single generator of our framework utilizes the histogram equalization of CECT and NECT images so that the generator to effectively learn the image contrast. The experimental results demonstrate that the proposed method provides synthetic NECT images from CECT with high quality than the original cycleGAN, which reduces not only the radiation exposure for patients but also computational cost for the vessel visualization.