A new deep learning model for predicting IMRT dose distributions for lung cancer with dose masks.

Purpose: 3D U-Net deep neural networks are widely used for predicting radiotherapy dose distributions. However, dose prediction for lung cancer IMRT is limited to conventional radiotherapy, with significant errors in predicting the intermediate and low-dose regions. Methods: We included a mixed dataset of conventional radiotherapy and simultaneous integrated boost (SIB) radiotherapy with various prescription schemes. In addition to inputting CT images and anatomical structures, we incorporated dose mask information to provide richer local low-dose details. We trained five models with varying numbers of dose masks to investigate their impact on dose prediction models. Results: The inclusion of dose masks led to significant improvements in prediction accuracy for both the PTV and OARs. In particular, the mean absolute error (MAE) of dosimetric metrics for most OARs fell below 2%, and voxel-wise MAE within each structure steadily decreased as more dose masks were supplied-most notably in low-dose regions. These results demonstrate that incorporating dose masks effectively enhances training efficiency and prediction stability. Among models receiving varying numbers of dose masks, the configuration with ten masks achieved the highest predictive accuracy. Conclusion: This study proposes a dose mask-assisted method for lung cancer IMRT dose prediction. It demonstrates high accuracy and robustness in clinical radiotherapy scenarios with various prescription schemes, including conventional radiotherapy and SIB. The inclusion of additional dose masks significantly improved model performance, with prediction accuracy increasing as the number of masks increased.