Click here to


Are you sure ?

Yes, do it No, cancel

Synthetic CT Generation Using Unpaired Images in a CycleGAN with Identity Loss

Z Sun1 , S Baek1 , S Yaddanapudi1 , J St-Aubin1*, University of Iowa, Iowa City, IA


(Tuesday, 7/16/2019) 10:00 AM - 10:30 AM

Room: Exhibit Hall | Forum 2

Purpose: To rapidly generate synthetic CT (sCTs) from T1-weighted MRI for the purpose of MRI-only treatment planning. We developed a cycle-consistent generative adversarial network (CycleGAN) with a novel identity-loss that accelerates convergence at early stages of training. The unsupervised learning architecture of the CycleGAN makes it ideally suited for training with unpaired images which is important in abdominal and thoracic regions where accurate image registration between CT and MRI may not be possible.

Methods: A CycleGAN model was designed to learn the mappings between CT and MRI using two sets of generator and discriminator networks. We include an identity-mapping term into our loss function (termed identity-loss) which enforces structural similarity between the generated sCT and the MRI. The CycleGAN model was trained using unpaired head and neck CT and MRI randomly sampled from a training set of eight patients. Two additional patients were used for evaluating the mean absolute error (MAE) between the sCT and the original CT.

Results: Training of the CycleGAN was performed on a single NVIDIA 1080Ti GPU with 11 GB memory taking ~6 mins per epoch. Preliminary results yielded MAE of 122+/-222 HU and 133+/-247 HU for test patient one and two, respectively, after 20 epochs with identity loss, and 154+/-264 HU and 162+/-283 HU, respectively, after 20 epochs without identity loss. The synthetic CT generation for each patient took less than one second using a trained network on an NVIDIA 1080Ti GPU.

Conclusion: A CycleGAN with identity-loss accelerated the training using unpaired T1-weighted MRI and CT data. The trained CycleGAN model generated a sCT in less than one second making it viable for MRI-guided radiotherapy. Future work to increase the accuracy of our CycleGAN model can be achieved through increased the number of training data, and the inclusion of additional consistency loss functions during training.


Image-guided Therapy, MR


IM- Dataset analysis/biomathematics: Machine learning

Contact Email