• Journal of the Korean Society of Radiology
• /
• v.18 no.6
• /
• pp.715-724
• /
• 2024

We aimed to build an MR-CT interconversion model based on structure-constraints Cycle-constraints Generative Adversarial Neural Networks (scCycleGANs). We used MDCT (Somatom Definition Flash CT, SIEMENS, Germany) and 3.0T MRI (Ingenia 3.0T CX MRI, PHILIPS, Netherlands) as our hardware equipment and Python (3.12.6) and PyTorch (2.4.0) as software. The study model was scCycleGAN. We acquired 2,871 and 2,436 brain CT and MR (T2WI) images of 87 patients, respectively, and for a total of 5,307 medical images, CT and MR images taken at the same level were classified through primary evaluation, and 364, 27, and 8 pairs of images were labeled as training, validation, and test data, respectively. Then, we applied hybrid objective function to the GAN model based on the basic APS frameworks to build the model, and the evaluation of the generated model was divided into quantitative and qualitative evaluation. The qualitative evaluation was conducted on 10 radiologists with more than 20 years of experience, and the quantitative evaluation was set as PSNR, IOU, SSIM, and MAE. The results of the qualitative evaluation showed that the percentage of 'positive responses', defined as a response of 'Neutral' or better, was 63% and 96% for the Synthesis CT (sCT) and Synthesis MR (sMR) groups, respectively, while the quantitative evaluation metrics PSNR, SSIM, and MAE achieved the initial target values for both groups. Our study can be used as basic guided research in the field of medical image conversion and synthesis. And further research and complementary studies are expected to solve problems such as model lightweighting to reduce the dose burden on patients and medical costs if applied to clinical environments.