基于DCRU-Net的有限视角放疗辐射致声3D在体剂量重建方法-《中国医学物理学杂志》

文章信息/Info

Title:: DCRU-Net-based three-dimensional in vivo dose reconstruction method for limited-view radiation-induced acoustic imaging

文章编号:: 1005-202X（2026）04-0421-09

作者:: 赵新新1; 汪新怡1; 杜胤达1; 陈博湧1; 周凌宏1; 李永宝2; 宋婷1; 1.南方医科大学生物医学工程学院，广东广州 510515； 2.中山大学肿瘤防治中心/华南肿瘤学国家重点实验室/肿瘤医学协同创新中心，广东广州 510060

Author(s):: ZHAO Xinxin1; WANG Xinyi1; DU Yinda1; CHEN Boyong1; ZHOU Linghong1; LI Yongbao2; SONG Ting1; 1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China 2. Sun Yat-sen University Cancer Center/State Key Laboratory of Oncology in South China/Collaborative Innovation Center for Cancer Medicine, Guangzhou 510060, China

关键词:: 放疗辐射致声成像; 3D在体剂量监测; 反投影重建; 有限视角; 深度学习

Keywords:: Keywords: radiation-induced acoustic imaging three-dimensional in vivo dose monitoring back-projection reconstruction limited-view deep learning

分类号:: R318；R811.1

DOI:: DOI:10.3969/j.issn.1005-202X.2026.04.001

文献标志码:: A

摘要:: 目的：针对有限视角下放疗辐射致声（RA）成像用于在线剂量监测时存在的图像畸变与伪影问题，提出一种基于深度级联残差U-Net（DCRU-Net）的3D在体剂量重建方法，以提升剂量监测精度。方法：构建深度级联网络DCRU-Net，通过分步优化实现有限视角下声学信号与剂量分布的复杂映射。第一级子网络以反投影（BP）重建的初始压力图为输入，旨在恢复缺失的声场信息，重建全视角3D压力场；第二级子网络在物理转换的基础上进一步校正误差并细化特征，生成高精度3D剂量图。基于80例前列腺癌患者的临床CT及计划剂量数据，利用k-Wave工具箱模拟会阴区2D传感器阵列采集的有限视角RA信号，并引入组织异质性、声速变化及噪声。采用相对均方根误差（rRMSE）、结构相似性（SSIM）及Gamma通过率作为主要评估指标。结果：定性分析显示，DCRU-Net能有效消除有限视角引起的失真与伪影，重建结果与标签高度一致。定量评估显示，预测压力图与剂量图的rRMSE分别为3.40%和2.50%，剂量图的SSIM达0.98。在3%/5 mm标准下，0%和70%阈值的Gamma通过率分别为98.18%和99.32%。此外，该方法在不同噪声水平下均表现出较强的稳定性。结论：本研究方案实现了高效率、高精度的有限视角3D剂量重建，克服了实际临床中传感器覆盖范围受限的问题，为RA成像在实时在体剂量监测中的应用提供了重要技术手段。

Abstract:: Abstract: Objective A novel three-dimensional (3D) in vivo dose reconstruction method based on a deep cascaded residual U-Net (DCRU-Net) is proposed to address the issues of image distortion and artifacts in limited-view radiation-induced acoustic (RA) imaging for online dose monitoring, thereby improving the accuracy of dose monitoring. Methods A DCRU-Net was constructed ?o achieve the complex mapping from limited-view acoustic signals to dose distribution through stepwise optimization. The first-stage sub-network took the initial pressure map reconstructed by back-projection (BP) as input, aiming to recover the missing acoustic information and reconstruct a full-view 3D pressure field. The second-stage sub-network further corrected errors and refined features based on physical conversion, generating a high-precision 3D dose map. Based on the clinical CT and planned dose data from 80 prostate cancer patients, the k-Wave toolbox was adopted to simulate limited-view RA signals acquired by a two-dimensional transducer array in the perineal region, with consideration of tissue heterogeneity, acoustic velocity variations, and noise. The relative root mean square error (rRMSE), structural similarity index (SSIM), and Gamma pass rate were taken as the primary evaluation metrics. Results Qualitative analysis demonstrated that the DCRU-Net effectively eliminated distortion and artifacts caused by limited-view acquisition, with reconstructed results showing high consistency with the ground truth maps. Quantitative evaluation revealed that the predicted pressure maps and dose maps achieved rRMSE of 3.40% and 2.50%, respectively, and the dose maps achieved an SSIM of 0.98. Under the 3%/5 mm criterion, the Gamma pass rates at 0% and 70% threshold conditions were 98.18% and 99.32%, respectively. Furthermore, the proposed method exhibited strong stability across different noise levels. Conclusion The proposed method realizes efficient and high-precision 3D dose reconstruction from limited-view data, and overcomes the practical clinical limitation of restricted transducer coverage, providing a significant technical approach for the application of RA imaging in real-time in vivo dose monitoring.

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

文章信息/Info

备注/Memo

常用功能

导航/Navigate

工具/Tools

统计/Statistics