Click here to


Are you sure ?

Yes, do it No, cancel

Performance Comparison of Conditional and Cycle-Consistent Generative Adversarial Networks Used for Abdominal Synthetic CT Generation

J Fu1*, K Singhrao1 , M Cao3 , Y Yang4 , D Ruan5 , J Lewis6 , (1) David Geffen School of Medicine at UCLA, Los Angeles, CA, (3) UCLA School of Medicine, Los Angeles, CA, (4) UCLA, Los Angeles, CA, (5) UCLA School of Medicine, Los Angeles, CA, (6) UCLA School of Medicine, Los Angeles, CA


(Sunday, 7/14/2019) 4:00 PM - 5:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: Deep learning methods have been used to achieve excellent accuracies in generating brain and pelvic synthetic CTs (sCTs) for MR-only radiotherapy. However, few such methods have been investigated for abdominal sCT generation. In this work, we compared two deep learning approaches, conditional generative adversarial network (cGAN) and cycle-consistent GAN (cycle-GAN), for abdominal sCT generation based on a single 0.35T MR scan.

Methods: CT and MR images of 12 abdominal cancer patients were studied retrospectively. N4 bias field correction and histogram-based normalization were performed on the MR images to minimize inter-patient intensity variations. All CTs were deformably registered to MR images to generate paired MR/deformed CT (dCT) for cGAN training. Since cycle-GAN does not require paired images for training, we investigated two training schemes: one using unpaired MR and dCT images, and the other using unpaired MR and non-deformed CT images. Model performance was evaluated by computing the mean absolute error (MAE) and peak signal to noise ratio (PSNR) between each sCT and dCT under a four-fold cross-validation protocol. Statistical tests were conducted to compare these metrics.

Results: The average MAEs were 88.0±15.7 HU, 99.1±17.8 HU, and 113.9±20.4 HU for cGAN, cycle-GAN trained with dCTs, and cycle-GAN trained with non-deformed CTs, respectively. The corresponding average PSNRs were 27.5±1.5, 26.8±1.3, and 25.7±1.4, respectively. Statistically significant differences were observed for all pairwise metric comparisons except for the PSNRs between cGAN and cycle-GAN trained with dCTs.

Conclusion: cGANs generated the most accurate abdominal sCTs. Cycle-GANs trained with dCTs performed better than cycle-GANs trained with non-deformed CTs. Since Cycle-GANs can be trained unsupervised, they may have access to training datasets larger than those available for cGANs, which may counterbalance the accuracy advantage of cGANs demonstrated in this 12-patient study. Further investigation is required to assess performance differences with larger training sets.

Funding Support, Disclosures, and Conflict of Interest: Varian master research agreement


Not Applicable / None Entered.


IM/TH- MRI in Radiation Therapy: MRI for treatment planning

Contact Email