Purpose: To implement a method for precise three-dimensional(3D) dose prediction in intensity-modulated radiation therapy(IMRT) using a dilated convolution deep convolutional neural network(CNN).
Methods: Seventy-seven cases of patients with cervical carcinoma were retrospectively selected. Sixth-six cases were chosen randomly as the training set and the remaining as the test set. A deep convolutional neural network model named Deep-ResUnet101 was trained to predict 3D dose distribution. First, the dose distributions and all contours were converted to 3D matrix and 3D binary matrix respectively using in-house developed python programs. Second, the contours and dose distributions at voxel-level were fed to the dilated neural network. The model was trained from scratch with weights randomly initialized. The trained deep convolutional network was used to predict 3D dose distributions, and the accuracy was evaluated against the corresponding clinical dose distribution.
Results: Our results demonstrate that the dilated neural network could predict 3D dose distribution precisely. For 11 tested cases, there was no significant difference between the predicted dose volume histograms (DVH) parameter values and the clinical ones. The mean absolute max dose differences and the mean absolute mean dose differences for PTV and OARs are within 2.2% and 2.8%, and nearly all the dosimetric indices for PTVs and OARs are within 0.1%. The average dice similarity coefficient (DSC) of different isodose volumes was 0.94-1.
Conclusion: This study developed a new dilated convolutional neural network for voxelized 3D dose prediction, which can precisely predict the 3D dose distribution for cervical carcinoma IMRT patients. The predicted 3D dose distribution could be used to improve the design of IMRT plan, ensure the plan quality and consistency, and guide automatic IMRT treatment planning.
Treatment Planning, Radiation Therapy, Dose
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation