H Li¹, F Zhou^1,2, B Liu^1,2*, S Xu³, Q Wu⁴, (1) Image Processing Center, Beihang University, Beijing, CN (2) Beijing Advanced Innovation Center for Biomedical Engineering, Beihang University, Beijing, CN (3) National Cancer Center/Cancer Hospital- Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, BJ, CN, (4) Duke University Medical Center, Durham, NC

• 
  H Li

PO-GePV-T-419 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Circular collimator is widely used for stereotactic radiosurgery on Cyberknife system. However, the vendor-supplied dose calculation algorithms suffer from the problem of either low accuracy in heterogeneous medium using RayTracing (RT) algorithm, or long computation time using Monte Carlo (MC). This work aims to develop a novel deep learning based dose engine which can combine the accuracy of MC and efficiency of RT.

Methods: A CNN network with encoder-decoder structure was proposed to predict dose with matching accuracy with MC. It takes the RT dose for a beam and the CT image as input, and outputs the predicted beam dose. A cohort of 10 lung cancer and 3 liver cancer patients were collected. The MC dose for each beam was simulated using the BEAMnrc/DOSExyz. The RT dose for each beam was calculated using an in-house GPU implementation. For the lung cases, five-fold cross validations were performed. In each fold, the beams of 8 patients were used for network training and the remaining 2 for testing. The trained model was applied on liver patient to test its generalization ability.

Results: With MC as reference, the average MSE was reduced from (2.1±1.8)×10⁻⁴ to (1.8±1.4)×10⁻⁵. The average gamma passing rate for the plan doses were improved from 83.42% to 98.31% at 2mm/2%. The average PTV V₁₀₀ and conformity index of the resulting dose (62.3%/1.09) were obviously closer to MC dose (61%/1.09) than RT (88.4%/1.36). The model had a good generalization ability and could improve the average gamma pass rate from 82.8% to 95.0% at 2mm/2% for the liver cases. The calculation time was less than 1 minute for a plan with 120 beams.

Conclusion: It is promising to use deep learning technique to calculate Cyberknife circular collimator treatment plans with both high accuracy and efficiency.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Key R&D Program of China under Grant No. 2018YFA0704100 and 2018YFA0704101.

Keywords

Dose, Monte Carlo, Radiation Therapy

Taxonomy

TH- External Beam- Photons: Computational dosimetry engines- Monte Carlo

Contact Email