Room: Exhibit Hall | Forum 5
Purpose: To make 3D dose predictions by analyzing the field geometry information together with patient specific images.
Methods: We propose a 3D-Convolutional Neural Network dose prediction model that processes 3D images together with field geometry. The networkâ€™s input has shape nX256X256Xc: n slices each with c channels of size 256X256 (the first (c-2) channels contain sets of organ masks, the last 2 channels contain CT images and the field geometry image). The OAR images are binary masks with value 1 for positions belonging to the OAR and 0 everywhere else. The OAR masks can be grouped together following the clinicâ€™s organ importance ranking. For the PTV masks, we use the scaled dose level value for the positions belonging to a given PTV and 0 everywhere else. The output of the network consists of a cube of shape (n-2)x256X256 representing the dose predictions for to the input image cube, except the first and last slices.
Results: We trained our model on a heterogeneous set of headâ€?andâ€?neck cancer cases with variability in tumor location and field geometry used. We compared the predicted dose distribution with the clinically approved dose distribution. The test results indicate the model learns to incorporate the field geometry information in the patient specific images and accurately predicts 3D dose maps for all treatment types.
Conclusion: Our model can be trained on heterogeneous datasets that exhibit variations in the location, size and shape of the PTVs which in turn lead to variations in the field geometry. Furthermore, our 3D dose prediction model could be trained using both coplanar and non-coplanar treatment plans. So, it could be used to make quick comparisons between 3D dose maps predicted with different field geometries and thus help the planner choose the most suitable field geometry for a given patient.
Funding Support, Disclosures, and Conflict of Interest: All authors of this work are employed by Varian Medical Systems.