Dose Prediction Using Three-Dimensional Convolutional Neural Network for Nasopharyngeal Carcinoma with Tomotherapy
Y Liu1,3*, G Zhang1, Z Chen2, J Wang3, X Wang3, B Qu3,4, S Xu3,4, (1) Beihang University, Beijing, BJ, CN, (2) Manteia Medical Technologies, Xiamen, CN, (3) PLA General Hospital, Beijing, BJ, CN, (4) Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Medicine and Engineering, Beihang University, Beijing, 100191, China; and Key Laboratory of Big Data-Based Precision Medicine (Beihang University), Ministry of Industry and Information Technology
Presentations
PO-GePV-T-292 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To accurately predict 3D dose distribution based on the patient-specific gap between organs at risk (OARs) and PTVs, and small OARs for nasopharyngeal carcinoma (NPC) with Tomotherapy.
Methods: 3D matrix for a deep learning model of dose prediction was trained to learn more about the relationship between dose distribution and patient anatomical information by increasing the height of training matrix. And 3D HD U-net was used as model. 127 NPC patients treated with Tomotherapy were enrolled in this study, including 84 for training and 43 for testing. The multi-channel 3D input matrix including CT and ROIs, dose array was used as the output matrix. A comparative experiment was done between the training matrix height of 48 (Model Ⅰ) and 16 (Model Ⅱ). Mean absolute error (MAE) was taken as the loss function. The models’ performance was evaluated by the comparison and statistical analysis.
Results: The corresponding errors were calculated by comparing the predicted dose with ground truth dose. The deviations among the mean and maximum dose of planning target volumes (PTVs) and OARs were within 4.37%. Error for max dose of optic nerves-L in Model Ⅰ were 4.05±3.49%, comparing with 7.11%±5.10% in Model Ⅱ(p=0.002). The gamma passing rates of body and PTV for 3%/3mm criteria were 76.5±10.7% and 83.8±5.4% in Model Ⅰ, comparing with 43.3±5.4% and 73.6±6.6% in Model Ⅱ(p<0.001). The prediction error of D95 for PTV was 0.48±0.33% in model Ⅰ, comparing with 2.72±1.28% in Model Ⅱ(p<0.001).
Conclusion: It is important for deep learning dose prediction model to learn the distance of OARs and PTV. We found that large height of training matrix could improve the accuracy of dose prediction. Our dose prediction methods provide a milestone to build automatic Tomotherapy planning.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by a grant from Medical big data and AI R&D project (No. 2019MBD-043).
ePosters
Keywords
Dose, Tomotherapy, Treatment Planning
Taxonomy
Contact Email