(Sunday, 7/29/2018) 3:30 PM - 4:00 PM
Room: Exhibit Hall | Forum 9
Purpose: This study investigates the extent of which a deep-learning model operating in 3D can capture the relationship between a set of objective function weights and the corresponding optimal dose-distribution. Solving this problem is an important step towards predicting dose-distributions associated with Pareto plans for novel patients.
Methods: For a given prostate case with three conflicting objectives and two constraints, we consider 34 different Pareto plans. As input to the model, we combine the binary representations of the available region of interests with the trade-off objective weights for each plan. As output to the model, we use the Pareto optimal dose-distributions. The model is based on a dilated dense-net with ReLU activations operating in 3D, and the training objective is composed of the mean of the absolute error relative to prescription.
Results: Our model is able to transform objective weights to dose distributions that spatially resemble the corresponding optimal dose-distributions well. Furthermore, we are able to reconstruct the 34 Pareto dose-distributions with a mean absolute error of less than 0.4% of prescription.
Conclusion: This study proposes a model for predicting dose distributions that is based on a fully convolutional network operating in 3D. We have shown that our model can be used to encode the transformation from objective function weights to Pareto optimal dose-distributions for one prostate case.
Funding Support, Disclosures, and Conflict of Interest: Research partially funded by RaySearch Laboratories.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
