MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Novel Adaptive Dose Prediction Neural Network for IMRT Treatment Planning Using Flexible Selection of Beam Parameters

J Chun1,2*, S Olberg1 , T Zhao1 , B Cai1 , H Li1 , I Park2 , J Kim2 , S Mutic1 , J Park1 , (1) Department of Radiation Oncology, Washington University in St. Louis, St Louis, MO 63110,(2) Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, South Korea

Presentations

(Sunday, 7/14/2019)  

Room: ePoster Forums

Purpose: The planning process of intensity-modulated radiation therapy (IMRT) requires iterative revision, varies with the planner’s experience, and becomes unintuitive as the complexity of plan objective increases. There have been trials of automatic dose prediction using deep learning in recent years, but experiments were limited to equiangular beam angles for a constant number of beams due to the rigidity of the networks. In this study, we propose a novel adaptive dose prediction neural network (aDPN) that predicts IMRT plan according to flexible selection of beam parameters at the planner’s discretion.

Methods: aDPN was trained and cross-validated with 80 prostate bed patients who received 66.6 Gy with 7 to 9 beams at varying beam angles. Automatic dose prediction was performed using a convolutional neural network consisting of two structures: 1) a U-Net that has been widely used with great success in medical applications, and 2) the atrous spatial pyramid pooling module that efficiently encodes multi-scale contextual information from a single low-level feature map. Features of the beams such as overlapped lengths between the target and pseudo-beam lines based on minimal user-selected prior information were used as input features. These features were labeled with dose distributions from clinically approved plans. The trained model was applied to an unseen, random subset of the datasets and evaluated.

Results: Our result demonstrate that aDPN can predict clinically acceptable dose distribution without major statistical difference between prediction and real clinical plan. 7-beam and 9-beam dose distributions are precisely generated by selecting the beam numbers and angles prior to evaluation. Dose distributions are predicted by the trained network with a throughput time of 0.05 sec/slice.

Conclusion: The proposed aDPN for dose prediction offers the ability to reduce the time for optimizing treatment beam parameters while allowing for flexibility in planning selections.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email