부산대학교 도서관

In this paper, we investigate the generation of Lumbar Spine synthetic CT (sCT) images based on MR images for MR-only spinal cord injury treatment and surgery planning. CT and MRI provide complementary information and are both important for spine treatment and surgery planning. However, the acquisition of images of two different modalities interrupts the clinical workflow, adds to health care cost and poses challenges in registering the images for analysis. Translating MR images to CT images would result in seamless correlation between images and also save patients from exposure to ionizing radiation due to a CT examination. Using a large clinical dataset of 800 patients, we showed that a cycle consistent generative adversarial network (CycleGAN) can be trained with the unpaired, unaligned MR and CT images of the Lumbar Spine in generating realistic synthetic CT images. The trained model was evaluated using the paired MR and CT images of 8 patients, the average Mean Absolute Error (MAE) was found to be 184 HU with a standard deviation of 24 HU.