Reconstruction of DRR-Like KV-DR Using CycleGAN-Based Image Synthesis for Intra- and Extracranial SRT/SRS
D Lee*1,2, S LEE3, H Cho1, (1) Yonsei University, Wonju, KR, (2) Samsung Medical Center, Seoul, KR (3) Boston Medical Center, Boston, MA
Presentations
SU-D-TRACK 4-5 (Sunday, 7/25/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: A kV digital radiograph (kV-DR) using ExacTrac® X-ray source (BrainLAB AG, Munich, Germany) undergoes degradation of image quality due to radiotherapy accessories equipped for patient immobilization. We proposed a deep learning model based on cycleGAN image synthesis to improve the image quality of ExacTrac kV-DR.
Methods: A total of 2574 paired kV-DRs and digitally reconstructed radiograph (DRR) were obtained from stereotactic radiotherapy (SRT/SRS) treatments for patients with brain or spine cancer. The DRR-like synthetic kV-DRs were generated from kV-DRs directly obtaining from ExacTrac using a deep learning-based image synthesis network. 2166 and 234 of kV-DR and DRR pairs were used for training for testing, and 92 and 82 pairs were used for testing from brain and spine cases, respectively. The images were normalized into [0, 1], and then the order of the paired data cases was shuffled to prevent the overfitting in network training. The mean absolute error (MAE), the root-mean-square error (RMSE), and the peak signal-to-noise ratio (PSNR) were used to quantify the image quality of the synthetic kV-DR.
Results: The synthetic kV-DR (MAE⁼0.26±0.08, RMSE⁼0.33±0.09, and PSNR⁼9.83±0.33 dB) showed a better image quality compared to ExacTrac kV-DRs (MAE⁼0.34±0.14, RMSE⁼0.39±0.14, and PSNR⁼8.62±3.07 dB). The average MAE, RMSE, and PSNR values were 0.21±0.06, 0.29±0.07, and 11.06±8.45 dB for the brain cases and 0.31±0.07, 0.39±0.08, and 9.83±2.42 dB for spine cases, respectively.
Conclusion: We successfully suppressed the artifacts in the kV-DRs originated from radiotherapy accessories and generated a DRR-like synthetic kV-DRs from ExacTrac based on deep learning network. The proposed model can improve the tumor localization accuracy for SRS/SRT radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2020R1F1A1075741).
Handouts
Keywords
Image Guidance, Digital Imaging, Image Processing
Taxonomy
IM- X-Ray: Machine learning, computer vision
Contact Email