Predicting Prostate VMAT 3D Radiation Doses of Continuously Varying Organ Dose Trade-Offs Using a Conditional Variational Autoencoder

Abstract

Predicting 3D radiation doses from planning structures is a promising method of knowledge-based treatment planning. However, most models are designed to predict only one 3D dose distribution per patient, based on historical organ dose trade-offs. To allow customizable plan generation, in this study, we aim to show the feasibility of dose prediction in which the degrees of organ dose trade-off could be explicitly specified. Specifically, the bladder vs. rectum dose trade-off in prostate cancer was investigated. In an IRB-approved study, we obtained imaging and structure contours for 167 patients with prostate cancer who received definitive radiotherapy. Training data was generated by automatically creating 3 different plans for each patient: while keeping target dose patterns constant, 1 base plan was generated with optimization objectives directly based on the output of a custom RapidPlan model prediction (S = 0), 1 plan with the goal to significantly lower bladder dose relative to the rectum (S = -1), and 1 plan with the goal to significantly lower rectum dose relative to the bladder (S = 1). This process was achieved by adjusting priority values during optimization. S is a scalar indicating the degree of bladder vs. rectum dose trade-off (higher S = higher dose to the bladder relative to the rectum). A conditional variational autoencoder (cVAE) was constructed as the generative model. Training, validation, and testing sets consist of 124, 10, and 33 patients, respectively. During training, the inputs to the model were 3D structure masks with voxel values modified based on S, and the output was the corresponding 3D dose. For model testing, we selected 7 equispaced values of S in the range [-1, +1] for each of the 33 test patients, generated the 3D doses for each S value (normalized to D2% = 110%), and calculated the differences of key dosimetric parameters (for S levels other than 0) compared to the predicted base plan (S = 0). The mean and standard deviations for these differences were reported. The cVAE model converged after training for 800 epochs. As the value of S increased from -1 to +1, the target coverage remained similar, while the doses to the bladder and rectum increased and decreased, respectively, as expected (Table 1). This pattern was also confirmed by qualitative examination of dose-volume histograms for additional S values. We demonstrated the feasibility of predicting 3D radiation dose distributions for prostate cancer where the degrees of organ dose trade-off could be explicitly specified.