Room: Karl Dean Ballroom A2
Purpose: During treatment planning, exploring tradeoffs between different treatment planning goals is a trial and error process and often time-consuming. In this work we use deep learning to develop a 3D dose distribution prediction algorithm based on patient CT and desired dose volume histogram (DVH).
Methods: In this work, we use a densely connected U-Net to predict 3D dose distribution from the patientâ€™s contours and the desired DVH curve. Each DVH curve is converted into a vector which is treated as a separate channel for the input of U-Net. Convolution and max pooling on contours extract features and reduce the dimension down to the size to match the DVH vector size. 97 prostate patients with 10 different IMRT plans for each patient were used for this work. The training dataset includes 77 patients while 20 patients were chosen as testing dataset.
Results: The predicted 3D dose distributions were very similar to the ground truth. The difference maps between predicted dose and truly delivered dose showed an insignificant difference, indicating the model can predict an accurate dose distribution not only for a particular patient anatomy but also for different desired DVH curves. The mean square error (MSE) is employed as our evaluation metric, the mean value of MSE between different plans within one patient is ~8.4%, and different patients have similar trend between minimum and maximum of MSE.
Conclusion: In our preliminary study, we have successfully used the DVH constraints together with patientâ€™s anatomy as input for a U-Net to predict 3D dose distributions.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by Cancer Prevention and Research Institute of Texas (CPRIT) (IIRA RP150485)