MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Variable-OAR Volumetric Dose Prediction Model for Radiation Therapy Using Deep Learning

m mashayekhi1*, I Tapia2, A Sadeghnejad Barkousaraie1, A Balagopal1, X Zhong1, S Jiang1, D Nguyen1, (1) The University of Texas Southwestern Medical Ctr, Dallas, TX, (2) University Of Texas At Dallas

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: treatment planning pipeline includes delineation of organs at risk, (OARs). The number of OARs to be contoured varies from patient to patient and according to the relevance and importance of OARs to the treatment site. Current dose prediction models are limited to fixed number of OARs as input. Thus either only important OARs need to be taken as input to the model, or the planners are required to delineate extra, in some cases unnecessary OARs. To overcome this problem and shorten the treatment planning process, we propose a dose prediction model to incorporate variable number of OAR into the dose prediction pipeline.

Methods: proposed model takes 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 patients, 54 training, 6 validation, and 10 test patients. Patient contours include PTV, body, bladder, rectum, left and right femoral heads. Each patient has 1200 generated Pareto surface intensity modulated radiation therapy (IMRT) plans. We used group normalization and means square error (MSE) loss function. The model was trained on 64,800 iterations, using the Adam optimizer with a learning rate of 1?10-3.


Results: We implemented a variable OAR deep neural network for 3D-dose prediction. To illustrate the applications of our model, we tested the performance of the model on prostate cancer patients. Comparison between true and predicted results showed promising performance of our model. As expected, by dropping one OAR the overall predicted dose distribution on other sites does not change significantly.


Conclusion: model is able to include physician preference into dose prediction process and moreover it in independent of the number of input OARs.

Funding Support, Disclosures, and Conflict of Interest: Varian

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email