Room: AAPM ePoster Library
This study investigates the benefit of incorporating explicit flexible constraints imposed on the dose volume histogram (DVH) when training a convolutional neural network (CNN) to predict deliverable dose distributions for automated prostate radiotherapy (RT) treatment planning.
The dataset contained 73 prostate cancer patients who were treated with VMAT using a prescription dose of 77Gy, of which 31 received a focal boost to 95Gy on the macroscopic tumor. A U-net regression CNN (CNN1) was trained to predict the 3D dose distribution from the planning CT and contours as input, using mean squared difference as loss function and 5-fold cross-validation. A second, identical CNN (CNN2) was trained similarly with an additional term in the loss function that directly compared the DVHs of the predicted and the ground truth dose distributions, thus incorporating domain specific knowledge considered during plan optimization. For one fold (13 patients), deliverable RT plans were generated in Varian Eclipse™ using the DVH of the predicted dose distributions as dose-volume objectives for plan optimization, i.e. dose mimicking. The performance of both CNNs was evaluated by the difference between predicted and ground truth dose distributions of clinically relevant dose-volume parameters (DVPs). These DVPs were also evaluated after dose mimicking to assess the impact of imposing deliverability.
For all DVPs, the mean absolute difference was reduced for prediction with CNN2 compared to CNN1 (PTV D95: 2.09% vs 3.21%; Dmax rectum: 1.52% vs 1.98%; bladder: 2.28% vs 2.73%). The difference in DVPs decreased between both CNNs after dose mimicking. However, the number of acceptable treatment plans was still larger for CNN2 than for CNN1 (10/13 vs 8/13).
Including the DVH explicitly during CNN training improved dose prediction performance. Although subsequent dose mimicking decreased differences between DVPs, a higher rate of acceptable plans was obtained by incorporating dose-volume constraints.