Three-dimensional dose prediction for lung IMRT patients with deep neural networks: robust learning from heterogeneous beam configurations

Abstract

Purpose: The use of neural networks to directly predict three-dimensional dose distributions for automatic planning is becoming popular. However, the existing methods use only patient anatomy as input and assume consistent beam configuration for all patients in the training database. The purpose of this work was to develop a more general model that considers variable beam configurations in addition to patient anatomy to achieve more comprehensive automatic planning with a potentially easier clinical implementation, without the need to train specific models for different beam settings.

Methods: The proposed anatomy and beam (AB) model is based on our newly developed deep learning architecture, and hierarchically densely connected U-Net (HD U-Net), which combines U-Net and DenseNet. The AB model contains 10 input channels: one for beam setup and the other 9 for anatomical information (PTV and organs). The beam setup information is represented by a 3D matrix of the non-modulated beam's eye view ray-tracing dose distribution. We used a set of images from 129 patients with lung cancer treated with IMRT with heterogeneous beam configurations (4-9 beams of various orientations) for training/validation (100 patients) and testing (29 patients). Mean squared error was used as the loss function. We evaluated the model's accuracy by comparing the mean dose, maximum dose, and other relevant dose-volume metrics for the predicted dose distribution against those of the clinically delivered dose distribution. Dice similarity coefficients were computed to address the spatial correspondence of the isodose volumes between the predicted and clinically delivered doses. The model was also compared with our previous work, the anatomy only (AO) model, which does not consider beam setup information and uses only 9 channels for anatomical information.

Results: The AB model outperformed the AO model, especially in the low and medium dose regions. In terms of dose-volume metrics, AB outperformed AO by about 1-2%. The largest improvement was found to be about 5% in lung volume receiving a dose of 5Gy or more (V₅ ). The improvement for spinal cord maximum dose was also important, that is, 3.6% for cross-validation and 2.6% for testing. The AB model achieved Dice scores for isodose volumes as much as 10% higher than the AO model in low and medium dose regions and about 2-5% higher in high dose regions.

Conclusions: The AO model, which does not use beam configuration as input, can still predict dose distributions with reasonable accuracy in high dose regions but introduces large errors in low and medium dose regions for IMRT cases with variable beam numbers and orientations. The proposed AB model outperforms the AO model substantially in low and medium dose regions, and slightly in high dose regions, by considering beam setup information through a cumulative non-modulated beam's eye view ray-tracing dose distribution. This new model represents a major step forward towards predicting 3D dose distributions in real clinical practices, where beam configuration could vary from patient to patient, from planner to planner, and from institution to institution.

Keywords: automatic planning; deep learning; dose prediction.