Exhibit Hall | Forum 7
Purpose: Seed implant brachytherapy (SIBT) is an important treatment modality for many kinds of cancers. However, the commonly-used dose calculation method based on TG-43 report (TG-43 method) often overestimates the dose, which may lead to suboptimal clinical dose distributions. The purpose of this study is to develop a deep convolutional neural network (DCNN) based dose engine which can generate accurate dose maps as good as Monte Carlo Simulations (MCS).
Methods: A cohort of 25 head and neck (H&N) patients who received SIBT was collected for method development and validation. The DCNN dose engine takes single seed dose distribution calculated by an improved TG-43 method with inhomogeneity correction, the CT image and seed position map as inputs and predicts the same dose map as the MCS dose map. Adversarial strategy was adopted to improve the performance. Five-fold cross validation was performed to assess the performance of the DCNN model and ablation studies were performed to investigate the design of model inputs and architecture.
Results: Using MCS as ground truth, the proposed DCNN dose engine reduced the absolute percentage error of target volume (TV) D90 from (9.4 ± 7.4)% for TG-43 method to (1.3 ± 1.0)%, resulting in an accuracy promotion of 86.2%. The mean absolute percentage error of doses inside TV was reduced from (8.9 ± 4.2)% to (1.8 ± 0.8)%. Ablation studies justified the choice of the model inputs and the adversarial strategy. It took about 1.2 s to generate a dose map with 100 seeds, in comparison with 3 h for MCS.
Conclusion: A DCNN based dose engine for SIBT was developed and validated on 25 H&N patient cases. The proposed DCNN dose engine could generate dose maps close to MCS in just 1~2 seconds. It improves over the standard TG-43 method in both inhomogeneity and inter-seed attenuation corrections.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Key R&D Program of China under Grant No. 2018YFA0704100 and 2018YFA0704101.