MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Prediction of Three-Dimensional Dose Distributions with Deep Learning for Automatic Treatment Planning of Scanned Proton Therapy

A Barragan Montero1*, M Huet-Dastarac1, S. Teruel-Rivas1, K Souris1, D Nguyen2, S Jiang2, J. A. Lee1, E Sterpin1, (1) UCLouvain, Brussels, BE, (2) UT Southwestern Medical Center, Dallas, TX

Presentations

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

Room: AAPM ePoster Library

Purpose: Deep neural networks (DNN) are becoming a popular tool for automatic treatment planning in radiation therapy, with several proof-of-concepts for VMAT and IMRT treatments. However, the feasibility of DNN dose prediction for proton therapy remains to be addressed. The present work is the first to investigate the use of DNN for proton dose prediction for head and neck (H&N) cancer.



Methods: The DNN model is based on the popular U-Net, which was modified to achieve a more efficient gradient back-propagation, by including densely connected layers like in the DenseNet architecture. We used different input channels for the binary masks from the targets (CTVlow, 50Gy - 54.25Gy and CTVhigh, 70Gy) and organs. A set of 62 H&N cancer patients treated with pencil beam scanning, with the same beam configuration (4 beams), was used for training (50 patients) and testing (12 patients). All plans were generated in RayStation v8a (RaySearch Laboratories, Sweden), using robust optimization with 4 mm for setup errors and 3% for range errors. The accuracy of the model was evaluated by comparing the mean dose and other DVH metrics for the predicted and real doses, as well as the Dice coefficient for the isodose lines.


Results: The average error on the mean dose for the test set was overall below 5% of the CTVhigh prescription dose (70Gy). Only three organs (larynx, parotid, and right submandibular gland) had an error above 5%. The Dice coefficient remained above 0.9 for all isodoses except for the 70%-80% isodoses, where it fell to 0.8. The training time was about 8h and the inference time was around 20 s using a RTX 2080ti GPU.



Conclusion: This work demonstrated the feasibility of using DNN to predict proton dose distributions, which can later serve for high quality automatic dose planning.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email