Purpose: Deep learning methods have been explored to accelerate speed and boost accuracy of dose calculation, which of great significance in improving radiotherapy efficiency, by converting fast AAA dose calculation results to much slower but more accurate AXB dose calculation results. In this study we aim to develop a new deep learning scheme, by incorporating ROI contours into the training structure, to improve the performance of dose conversion from AAA dose to AXB dose.
Methods: We developed a 3D HD U-NET to convert AAA dose to AXB dose. In this deep learning scheme ROI contour binary masks were able to be incorporated as input data in the network structure, as well as the CT and AAA dose. Corresponding AXB dose were calculated using AXB algorithm as the ground-truth. 95 VMAT female pelvic cancer cases were selected for the research, where 76 cases were for training, and 19 cases were for testing. MSE is the loss function. Decay learning rate ranged from 10-4 to 10-6. The training process is stopped when the loss and learning rate stopped to decrease. The comparison study used only CT and AAA dose as input data.
Results: We calculated the Gamma passing rate (1 mm / 1%, 10% threshold) to evaluate our method using corresponding AXB dose calculation as comparison reference. The Gamma passing rate of original AAA dose and converted AXB dose are 73.64 ± 3.46% and 95.57 ± 1.06%, respectively. Time consumed for the conversion is 3.06 ± 1.78s. Compared to deep learning schemes using only CT and AAA dose as input data, Gamma passing rate of our newly devised method increased by 1.1%(TTest, p=0.004).
Conclusion: Deep learning method performs well in converting AAA dose to AXB dose in VMAT female pelvic cancer cases. Adding ROI contour as training data can improve the accuracy of dose conversion.
Not Applicable / None Entered.