(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 2
Purpose: To explore the feasibility of synthesizing CT image from magnetic resonance (MR) image using generative adversarial networks for nasopharyngeal carcinoma (NPC) intensity-modulated radiotherapy treatment (IMRT) planning.
Methods: T1-weighted MR image and CT images were acquired for each of 157 NPC patients. MR-CT images of 12 patients were randomly chosen as test set and the remaining images were used to build the conditional-GAN and cycle-GAN models. The conditional-GAN was trained with well registered MR-CT paired images and the cycle-GAN was trained with unregistered MR and CT images. For each tested patient, the synthetic CT (SCT) images were generated by these two models respectively from the MR image and were compared with the true CT images using the mean absolute error (MAE). The dosimetric accuracy of applying SCT images in dose calculations were evaluated on NPC IMRT plans.
Results: The computation time for generating an SCT volume image was 2.19 - 5.42 seconds for the two models. The MAEs within body were 68.23 ± 11.74 HU and 98.02 ± 7.69 HU for the conditional-GAN and cycle-GAN. The 2%/2mm ? passing rates were (98.59 ± 1.05) % and (98.13 ± 1.24) % for the conditional-GAN and cycle-GAN. The mean of absolute point dose discrepancies within regions of interest were (0.53 ± 0.25) % and (0.70 ± 0.41) % for the conditional-GAN and cycle-GAN.
Conclusion: The GAN models can generate accurate SCT images from MR image within seconds with high dosimetric accuracy for NPC IMRT planning. The conditional-GAN was preferred if high quality MR-CT image pairs were available.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
