Purpose: To develop an advanced deep convolutional neural network architecture to generate synthetic CT (SCT) images from MR images for intensity-modulated proton therapy (IMPT) treatment planning of head and neck cancer patients.
Methods: T1-weighted MR images and paired CT (PCT) images were obtained from 189 nasopharyngeal cancer (NPC) patients under radiotherapy immobilization. Deformable image registration was performed between MR and PCT images for each patient to create an MR-CT image pair. Thirteen pairs were randomly chosen as independent test sets and the remaining 176 pairs (14 for validation and 162 for training) were used to build two conditional generative adversarial networks (GAN)s: 1) GAN3D using a 3-dimensional Unet enhanced with residual connection and attentional mechanism as the generator and 2) GAN2D using a standard 23-layer 2-dimensional Unet as the generator. For each test patient, SCTs were generated using the generators with the MR image as input and were compared with respect to the corresponding PCT. A 4-beam IMPT treatment plan was created and optimized on the PCT, and then the dose matrix was recalculated on the SCTs. The dosimetric accuracy was evaluated using gamma index.
Results: The mean absolute error (MAE) between the PCT and SCT images, within the whole body, was (65.31±5.16)HU and (66.12±4.51)HU for GAN3D and GAN2D models, respectively. The MAE, within the bony structure (HU>150) were (172.01±16.71)HU and (175.7±18.54)HU for GAN3D and GAN2D models, respectively. The (2m/2% and 3m/3%) gamma passing rates were (98.2±1.7)% and (99.4±0.8)% for the GAN3D model and (97.7±1.9)% and (99.3±0.8)% for the GAN2D model.
Conclusion: SCT image generated using the conditional GAN achieved clinical acceptable dosimetric accuracy for IMPT plan of NPC patients. Using advanced network architecture design, such as residual connection and attention mechanism for the generator, SCT image can be further improved and resulted in a small improvement of dosimetric accuracy.