A method to establish reference benchmarks for in vivo dose monitoring for radiotherapy based on dual-energy cone beam CT and deep learning
10.3760/cma.j.cn112271-20240824-00319
- VernacularTitle:基于双能锥束CT和深度学习的放疗在体剂量监测基准合成方法
- Author:
Huimin HU
1
;
Zhengkun DONG
;
Shutong YU
;
Chen LIN
;
Tian LI
;
Yibao ZHANG
Author Information
1. 北京大学医学部医学技术研究院，北京 100191
- Publication Type:Journal Article
- Keywords:
Online adaptive radiotherapy;
In vivo dose monitoring;
Dual-energy cone-beam computed temography;
Generative adversarial network (GAN)
- From:
Chinese Journal of Radiological Medicine and Protection
2025;45(2):129-136
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To achieve the conversion from dual-energy cone-beam CT (DECBCT) at the kilovolt (KV) level to projections at the megavolt (MV) level using an improved CycleGAN network, in order to provide a potential reference benchmark and real-time monitoring of in vivo doses delivered by exit beams for the safe implementation of advanced techniques such as online adaptive radiotherapy. Methods:Simulated patient data were generated using a 4D extended cardiac torso (XCAT) model, and projections were generated based on the geometric parameters of Varian′s onboard cone-beam CT. Furthermore, relative electron density (RED) images were derived from DECBCT images using an iterative dual-energy decomposition algorithm. The SE-CycleGAN and CycleGAN networks were trained to generate MV projection images using DECBCT projections and RED images, respectively. The performance of both methods was evaluated using metrics including structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and root mean square error (RMSE).Results:SE-CycleGAN significantly outperformed CycleGAN in all evaluation metrics ( Z = -23.92, -26.17, -25.54, -26.80, -11.54, -11.21, P<0.05), particularly in learning global information. Besides, although both methods generated satisfactory MV projections, training using DECBCT projections as input yielded better effects than training using RED images. For all the 3 636 sets of projections in the test set, the SE-CycleGAN and CycleGAN networks using DECBCT projections as input respectively yielded SSIMs of 0.997 7±0.000 7 and 0.997 1±0.001 6, PSNRs of 39.625 0±4.684 4 and 36.272 2±5.566 3, and RMSEs of 0.004 1±0.002 7 and 0.006 3±0.0043, respectively. In contrast, the SE-CycleGAN and CycleGAN networks using RED projections as input respectively yielded SSIMs of 0.996 8±0.001 0 and 0.996 2±0.001 5, PSNRs of 38.548 7±3.637 4 and 36.007 3±4.437 8, and RMSEs of 0.004 3±0.002 2 and 0.006 1±0.0037, respectively. Conclusions:This study proposed a new method to establish reference benchmarks for in vivo dose monitoring based on DECBCT and deep learning technologies. This method is accurate and effective according to the preliminary validation using virtual simulation experiments.
