Room: Room 207
Purpose: To learn treatment plan models that accurately predict voxel-wise doses from anatomy. Convolutional neural networks (CNNs) learn data features and emulate the non-linear mappings common to RT planning. Such models could drive more accurate planning and adaptive plan/re-plan decision-making.
Methods: CNNs of two types learned voxel-wise mappings of 3D anatomies to corresponding 3D dose distributions. Axial binary masks or signed distance maps of the target PTVs and nearby OARs were aligned with axial doses and presented as paired 2D images for training and testing. 178 prostate cases were planned using Erasmus iCycle fluences and Monaco-segmented doses (ground truth plans). 165 cases were used for training; 13 cases for testing. Learning was done using U-Net on Caffe and conditional generative adversarial networks (cGAN) on TensorFlow. Testing against the ground truth iCycle/Monaco plans was done at training intervals.
Results: DVH differences served as the main CNN performance metric. dDVH is the mean absolute dose (cGy) difference per volume%, averaged over the 13 test estimate doses versus the corresponding ground truth doses. The minimum PTV dDVH (using a cGAN) was 24 cGy /volume%-bin and exhibited tight agreement at the high dose and low volume% DVH extrema. OAR DVHs showed only small differences throughout training. The U-Net results were uniformly poorer. In all experiments the dDVH descends to a minimum during training then rises as training proceeds, clear evidence of overfitting. Also, the details of DVH curve matches, and the detailed differences in 3D dose difference maps show complex, non-monotonic trends throughout training.
Conclusion: This project demonstrated CNN capability to model iCycle/Monaco prostate treatment plans and produce small DVH differences between test dose estimates and ground truth plans. As important, this project explored the details of training to illuminate the complexities of CNN learning and inference.
Not Applicable / None Entered.