Y Nomura¹*, J Wang^2,3 , H Shirato^2,3 , S Shimizu^2,4 , L Xing^2,5 , (1) Department of Radiation Oncology, Graduate School of Medicine, Hokkaido University, Sapporo, Japan, (2) Global Station for Quantum Medical Science and Engineering, Global Institution for Collaborative Research and Education (GI-CoRE), Hokkaido University, Sapporo, Japan, (3) Department of Radiation Medicine, Faculty of Medicine and Graduate School of Medicine, Hokkaido University, Sapporo, Japan, (4) Department of Radiation Medical Science and Engineering, Faculty of Medicine, Hokkaido University, Sapporo, Japan, (5) Department of Radiation Oncology, Stanford University, Stanford, CA

• 
  Y Nomura

(Tuesday, 7/16/2019) 7:30 AM - 9:30 AM

Room: Stars at Night Ballroom 1

Purpose: This study proposed a nearly-real-time spot-scanning proton dose calculation method using a three-dimensional convolutional neural network (3D-CNN).

Methods: First, CT images and clinical target volume (CTV) contours of total 215 head and neck patients were collected from TCIA database. 700 CTVs were extracted for training of the deep learning-based dose calculation model, while 28 CTVs were used to test trained model. Second, the optimal spot beam sequence data and dose distribution were calculated for each case using a Matlab-based dose calculation toolkit named matRad. The variable spot data were then converted into a fixed size of volume called peak maps (PMs). Third, a Unet-based 3D-CNN, which includes deformable pooling layers, was generated. Input and label of the 3D-CNN were set to a stacked volume of proton stopping power ratio images and the PMs and the dose distribution, respectively. Total 150-epoch end-to-end training was implemented with data augmentations of random volume flip and 90-degree rotations. Finally, accuracy of the 3D-CNN-calculated dose distributions were evaluated with the true distributions using a few evaluation metrics.

Results: The 3D-CNN model calculates proton dose distributions accurately with an average error of -0.0288 GyE and standard deviation of 0.116 GyE. Relatively high dose errors over 1 GyE were observed at body surface and high contrast edges. The 3D-CNN-calculated dose distributions have similar CTV dose-volume histograms to the ground truths. Averaged Dice similarity coefficients between binarised output and label distributions are mostly higher than 80%. Computation time for calculating one dose distribution is around 0.12 seconds with a NVIDIA GTX 1070 GPU.

Conclusion: A novel spot-scanning proton dose calculation method using 3D-CNN was developed. Our deep learning model is able to calculate 3D dose distributions with any spot data and beam irradiation angles and can be utilized for dose verification and image-guided proton therapy.

Keywords

Protons, Dose, 3D

Taxonomy

TH- External Beam- Particle therapy: Proton therapy - computational dosimetry-deterministic

Contact Email