Toward Online Adaptive Proton Therapy of Sinonasal Cancer: The Dosimetric Accuracy of Utilizing Deep- Learning Synthetic CT From CBCT for Daily Treatment Dose Verification
The American Association of Physicists in Medicine
Purpose: Proton beam is sensitive to anatomy variation, especially for sinonasal cancer patients due to daily nasal cavity fillings variation and tumor shrinkage. With online CBCT increasingly available and daily acquired in the proton treatment room, we hypothesized that daily proton dose can be accurately reconstructed from CBCT using a deep-learning model for sinonasal cancer patients.
Methods: 35 pairs of pre-treatment CT and first-fraction CBCT were selected from patients treated with proton therapy in our institution for training. Deformable-registration(DIR) was performed to eliminate potential anatomy variation between the simulation and first treatment. A Generative-Adversarial-Networks(GAN) model, including one generator with Residual-Unet architecture, and one discriminator using a multi-layer convolutional neural net(CNN), were trained from scratch to predict CT from CBCT, with random 3D B-spline deformation as augmentation. Five additional sinonasal patients with same-day re-plan CT and CBCT were selected for dosimetric evaluation. The re-plan CTs were deformed to the same-day CBCT to eliminate any neck flexion variation and served as the ground-truth(GT-CT). The synthetic CTs(Syn-CT) were compared to the GT-CTs by Mean-Absolute-Error(MAE). The clinical plans(3-5beams) were re-calculated on both Syn-CT and GT-CT using a commercial Monte-Carlo algorithm with 0.5% uncertainty. Clinically relevant DVH parameters and γ-passing rates were utilized to quantify the dosimetric discrepancy.
Results: The prediction takes less than 10s per patient. The MAE of Syn-CT was 53.58±5.3HU. Compared to the reference doses on GT-CT, the average1%/1mm,2%/2mm,3%/3mm γ-passing rates of the doses on Syn-CT were 82.6%,97.5%,and99.0%. The mean D_99 of CTV,D_1 of (brain-stem,cord,optical nerves,chiasm), and D_mean of parotids on Syn-CT are 99.9%,101.7%,102.5,104.5%,98.9%,and 100.2% of those on GT-CT, respectively.
Conclusion: The Synthetic CTs predicted by the GAN-model achieved clinically acceptable dosimetric accuracy as evaluated by the same-day re-plan CT. The method could be utilized to monitor daily treatment dose online and trigger plan adaptation for patients with undesired dosimetric variations.
A Qin, L Zhao, S Chen, X Li, W Zheng, D Yan, R Deraniyagala, X Ding. Toward Online Adaptive Proton Therapy of Sinonasal Cancer: The Dosimetric Accuracy of Utilizing Deep- Learning Synthetic CT From CBCT for Daily Treatment Dose Verification. ; 2021. Available from: https://w4.aapm.org/meetings/2021AM/programInfo/programAbs.php?sid=9289&aid=57492.