An OAR Sparing Preference Prior Guided Lung RT Plan Dose Prediction Model
Abstract
Purpose
We developed an efficient mathematical representation of achievable Pareto optimal OAR dose sparing objectives for lung cancer RT planning. We further incorporated these preferences in a RT planning model for individualized RT dose prediction tailored to patient specific OAR sparing need.
Methods
A vanilla model to predict plan dose distribution from patient CT and structure contours was first trained to separate the plan dosimetrics variations due to patient anatomical variability from those due to OAR sparing preferences. A non-dominant sorting algorithm was used to order the deviations of the predicted dosimetric
Objective
Esophagus_V60, Heart_V50, and Lung_V20 from the actual plan values based on Pareto optimality and to extract the non-dominant objective subset. A denoising auto-encoder (DAE) was then employed to extract a 2D compressed latent representation of the sparsely populated Pareto set by utilizing its bottleneck architecture. It was then used to augment the dose prediction model inputs. VMAT lung RT plans for 98 patients were included in this study with an 80/20 training/testing split ratio. The quality of the 2D latent space representation of the Pareto surface (PS) and the prediction accuracy improvement by the augmented model was assessed.
Results
The non-dominant subset of predicted dosimetric deviations from the actual plans shows a wide range of different sparing preferences and strong non-convex behavior. The Mean Squared Distances of the dosimetric deviations to the latent space representation is 1.56 (in % of organ volume). The Mean absolute error (MAE) of the predicted 3D dose distribution by the augmented model is 3.6 Gy, compared with the MAE of 4.4 Gy from the vanilla model.
Conclusion
The DAE provides efficient compressed representations for non-convex PS of RT plan DVH metrics. The augmented model can better account for the clinical trade-off decisions in clinical data and can provide individualized dose prediction.