Purpose: To develop an artificial intelligence―deep learning-based method for fully automated, accurate, and rapid prediction of optimized head-and-neck 3D dose distributions.
Methods: The DeepDoseNet 3D dose prediction model is based on ResNet and Dilated DenseNet. The 340 head-and-neck datasets from the 2020 AAPM OpenKBP challenge were utilized, with 200 for training, 40 for validation, and 100 for testing. Structures include 56Gy, 63Gy, 70Gy PTVs, and brainstem, spinal cord, right parotid, left parotid, larynx, esophagus, and mandible OARs. Mean squared error (MSE) loss, mean absolute error (MAE) loss, and MAE plus domain-knowledge (DK)-based loss functions were investigated. Each model’s performance was compared using a dose score, ¯S(D), (mean absolute difference between ground truth and predicted 3D dose distributions) and a DVH score, ¯S(DVH) (mean absolute difference between ground truth and predicted dose-volume metrics).
Results: DeepDoseNet requires ~0.5 sec to predict the 3D dose distribution on NVIDIA Quadro RTX 8000 GPU with 48GB memory. The MAE+DK model had the lowest prediction error (P<0.0001, Wilcoxon test) on validation and test datasets (validation: ¯S(D)=2.3Gy, ¯S(DVH)=1.9Gy; test: ¯(S_D )=2.0Gy, ¯S(DVH)=1.6Gy) followed by the MAE model (validation: ¯S(D)=3.6Gy, ¯S(DVH)=2.4Gy; test: ¯S(D)=3.5Gy, ¯S(DVH)=2.3Gy). The MSE model had the highest prediction error (validation: ¯S(D)=3.7Gy, ¯S(DVH)=3.2Gy; test: ¯S(D)=3.6Gy, ¯S(DVH)=3.0Gy).
Conclusion: The addition of the domain-knowledge reduced ¯S(D) by ~60% and ¯S(DVH) by ~70%. DeepDoseNet’s optimized dose can be used to improve the efficiency of dose optimization algorithms by providing a good estimate of the final dose distribution. Because of DeepDoseNet’s rapid and fully automated execution, it can be used in automated workflows to estimate the dosimetric relevance of delineated organs-at-risk (OARs).
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01CA222216