Synthetic Enhancement of Cone Beam Computed Tomography (CBCT) for Adaptive Radiotherapy Using Deep Learning Algorithms

Abstract

Materials/Methods: Thirty-three and 30 pairs of pre-treatment CT and first-fraction CBCT were selected from patients treated with proton therapy in our institution for Lung and H&N cancer, respectively. Deformable registration was performed to eliminate any potential setup variation between CT simulation and first fraction. The paired CT-CBCT datasets were then divided into training, validation and testing groups (lung patients: 24, 3 and 6, H&N patients: 22, 3 and 5 respectively). Two types of GAN models, including one generator with Orginal-Unet (Org-GAN) or Residual-Unet (Res-GAN) architecture, and one discriminator using a multi-layer convolutional neural net (CNN), were trained from scratch to predict CT from CBCT, with random deformation as data augmentation. Two training schemes were also tested, with two groups of patients trained separately or mixed together. All evaluations were performed on the testing group. The model-predicted synthetic CTs (synCT) were compared to the paired CTs in terms of Mean Absolute Error (MAE). The clinical treatment plans were recalculated on both synCT and paired CT using a commercial Monte-Carlo algorithm. Clinically relevant DVH parameters and g passing rates were utilized to quantify the dosimetric discrepancy. Results: The best MAE was achieved for Res-GAN model with the separately training scheme, 49.2AE6.7 and 49.6AE2.1 HU for Lung and H&N group, respectively. Compared to the reference doses calculated on the paired CT, the 3%/3mm and 2%/2mm g passing rates of the doses on Res-GAN synCT were 99.5AE0.4% and 95.3AE1.3 % for lung patients and 99.6AE0.4% and 97.0AE1.2% for H&N patients. The clinic relevant DVH parameters were within 1.0% for lung patients, except for cord D_1cc (1.9AE1.1%), mean dose of esophagus (1.3AE0.7%) and Trachea (1.1AE0.3%). For H&N patients, all DVH parameters agreed well within 0.5%, except for the mean dose of oral-cavity (2.7AE3.0%) and esophagus (2.3AE3.0%). Conclusion: With a relatively small number of training patients, the proposed Res-GAN model can predict synCT via CBCT with clinically acceptable dosimetric accuracy for proton treatment. The proposed method could be utilized to monitor potential proton dose deviation and trigger plan adaptation for lung and H&N patients.