A Site Agnostic 3D Dose Distribution Prediction Model for Radiation Therapy with Deep Neural Networks
m mashayekhi1*,A Balagopal1, I Tapia2, X Zhong1,S Jiang1, D Nguyen1, (1) The University of Texas Southwestern Medical Ctr, Dallas, TX, (2) University Of Texas At Dallas
Presentations
PO-GePV-M-235 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Radiotherapy data is hard to collect. Normally we can only get a couple hundred patient’s data to train the model on. Current dose prediction models are site specific, and need to be trained on the limited data available, this can reduce the overall potential performance. We propose a site agnostic model that can leverage data from any treatment site thus increasing total data available to train the model with.
Methods: The input to our proposed model is the patient’s 3D anatomical information and physician imposed tradeoff preferences to predict the 3D-dose prediction. Our supervised convolutional neural network, is based on 3D-U-net model. Data preprocessing includes mapping dose to voxels based on their distance ranks to PTV. The dataset contains 70 prostate cancer IMRT plans (source data) and 38 head and neck cancer VMAT plans (target data). We tested the generalizability of the model to different treatment sites via transfer learning technique. We have tested the performance of model with quantitative dose-volume-histogram (DVH) metrics for planning target volume (PTV) and organs-at-risk (OARs). We also tested the dependence of the model performance on the training data size to ensure the efficiency gain.
Results: We implemented a deep neural network for volumetric dose prediction that is independent of the number of OARs and the treatment site. We tested the performance benefit by leveraging data from 2 different sites: prostate and head and neck. The model trained on one treatment site, efficiently performed high quality predictions on the second site even with small training sizes.
Conclusion: Our proposed model is independent of the number of input OARs and the treatment site. Our model is able to make accurate predictions with limited training data.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email