Click here to


Are you sure ?

Yes, do it No, cancel

Three-Dimensional Dose Prediction with Deep Learning for IMRT Treatments of Variable Beam Configuration

A Barragan Montero1, 2*, D Nguyen2 , W Lu2 , M Lin2 , R Norouzi-Kandalan2,X Geets1, 3 , E Sterpin1, 4 , S Jiang2 , (1) UCLouvain, Center of Molecular Imaging, Radiotherapy and Oncology (MIRO), Brussels, Belgium, (2) UT Southwestern Medical Center, Medical Artificial Intelligence and Automation Laboratory (MAIA), Dallas-TX, USA (3) Cliniques universitaires Saint-Luc, Department of Radiation Oncology, Brussels, Belgium, (4) KU Leuven, Department of Oncology, Laboratory of Experimental Radiotherapy, Leuven, Belgium.


(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 1

Purpose: The use of deep neural networks to learn from previous clinical cases to predict the optimal dose distribution for new patients is becoming popular. However, the existing models are trained with a specific beam-setup, hampering their use for different beam configurations. This works aims to develop a general model that is robust to variable beam configuration, and therefore, suitable for realistic clinical environment.

Methods: The proposed Antatomy and Beam (AB) model combines U-Net and DenseNet architectures, and contains 10 input channels, one for beam-setup and 9 for anatomical information (PTV and organs). The beam-setup information is represented in the dose domain, by a 3D matrix of the non-modulated beam’s-eye-view ray-tracing dose. The AB model was compared with our previous work, the Anatomy Only (AO) model, containing only 9 channels for anatomical information. A set of 129 IMRT lung cases, treated with very heterogeneous beam configuration (4 to 9 beams of various orientations), was used for training (100 patients) and testing (29 patients). The model accuracy was evaluated by computing the absolute error of the predicted doses with respect to the clinical ones, as well as dice coefficients of their isodose volumes.

Results: The AB model outperformed the AO model, improving the mean absolute error of the predicted with respect to the clinical doses in about 1%-2% of the prescription dose, for relevant DVH metrics. Dice coefficients showed a 10% improvement in the low to medium dose region for the AB model with respect to the AO. The average prediction time was about 11 seconds for both models.

Conclusion: The AB model achieved accurate and fast dose prediction, while being robust to variable beam configuration. This opens the door to more comprehensive automatic planning with potentially easier clinical implementation, without the need of training specific models for different beam settings.

Funding Support, Disclosures, and Conflict of Interest: A Barragan Montero is supported by Baillet Latour Funds


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email