| 王伊玲,赵越,刘秋涵,等.基于人工智能的磁共振引导放疗靶区目标自动追踪研究[J].中华放射医学与防护杂志,2025,45(6):558-565.Wang Yiling,Zhao Yue,Liu Qiuhan,et al.Automatic target volume tracking in magnetic resonance imaging-guided radiotherapy based on artificial intelligence[J].Chin J Radiol Med Prot,2025,45(6):558-565 |
| 基于人工智能的磁共振引导放疗靶区目标自动追踪研究 |
| Automatic target volume tracking in magnetic resonance imaging-guided radiotherapy based on artificial intelligence |
| 投稿时间：2024-07-11 |
| DOI：10.3760/cma.j.cn112271-20240711-00257 |
| 中文关键词: 磁共振引导放疗 深度学习 电影磁共振图像 靶区追踪 |
| 英文关键词:Magnetic resonance image-guided radiotherapy Deep learning Cine magnetic resonance image Target volume tracing |
| 基金项目:四川省科技计划资助项目(2023YFH0079,2023NSFSC0720);国家自然科学基金(61901087) |
|
| 摘要点击次数: 5 |
| 全文下载次数: 1 |
| 中文摘要: |
| 目的 探讨Elekta Unity磁共振引导放疗系统靶区自动追踪的可行性,引入Transformer大模型深度学习技术,进一步提升磁共振引导放疗实时靶区目标追踪性能。方法 回顾性收集75位已接受Unity磁共振放疗的胸腹部恶性肿瘤患者4 661帧电影磁共振图像(cine MRI)二进制图像作为训练集;并额外收集10位患者500帧cine MRI二进制图像作为独立测试集。开发医学图像格式转化运算模块,将二进制图像转化为医学元图像。对测试集cine MRI实施肿瘤靶区外轮廓人工勾画,作为真实对照标签。固定每一位患者的第一帧cine MRI为参考图像,基于Transformer技术,构建运动图像(除第一帧之外的其余cine MRI图像)相对于参考图像的形变位移矢量场(DVF)深度学习模型。计算戴斯相似系数(DSC),95%豪斯多夫距离(HD95),雅格比行列式(NegJ),平均单帧cine MRI图像处理时间,并与传统B-Spline方案比较,定量评估靶区目标追踪准确性、DVF物理合理性及模型执行效率。结果 与B-Spline方案相比,所提出的深度学习方案具有更佳的靶区目标追踪性能,DSC [(0.84±0.05)vs.(0.74±0.16),t=11.44,P<0.05 ]和HD95 [(9.25±2.98)vs.(14.70±8.55)mm,t=-11.83,P<0.05 ] 均有改善;且图像平均处理时间由1.95 s缩减到30.99 ms,效率提升了2个数量级。深度学习方案获得了与B-Spline方案类似的NegJ,说明提取到的DVF具有与传统方案相似的物理合理性。结论 Transformer深度学习靶区自动追踪方案填补了Elekta Unity磁共振引导放疗系统的功能空白,可对胸腹部运动肿瘤目标实施较为准确且高效的自动追踪。 |
| 英文摘要: |
| Objective To explore the feasibility of automatic target volume tracing in the Elekta Unity magnetic resonance imaging (MRI)-guided radiotherapy system and to further enhance the real-time target volume tracing performance of MRI-guided radiotherapy by introducing the deep learning technology based on a large Transformer model. Methods A total of 4 661 frames of cine MRI binary images from 75 patients with malignant tumors in the chest/abdomen who were treated with MRI-guided radiotherapy were retrospectively collected as a training set. Another 500 frames of cine MRI binary images from 10 patients were collected as an independent test set. A module for medical image format conversion was developed to convert binary images into medical meta-images. The outer contours of tumor target volumes in the cine MRI images of the test set were manually delineated as actual control labels. With the first frame of the cine MRI images of each patient as the reference image and the other frames as motion images, a Transformer-based deep learning model was constructed to describe the deformable vector field (DVF) of motion images relative to the reference image. The Dice similarity coefficient (DSC), the 95% Hausdorff distance (HD95), the negative Jacobian determinant (NegJ), and the average processing time per frame of cine MRI images were calculated. These values were compared to those of the conventional B-Spline scheme to quantitatively assess the target volume tracing accuracy, DVF physical plausibility, and execution efficiency of the Transformer-based deep learning model. Results The Transformer-based deep learning model constructed in this study delivered superior target volume tracing performance, with improved DSC [(0.84 ± 0.05) vs.(0.74 ± 0.16), t = 11.44, P < 0.05] and HD95 [(9.25 ± 2.98) vs.(14.70 ± 8.55) mm, t= -11.83, P < 0.05]. Furthermore, this model reduced the average image processing time from 1.95 s to 30.99 ms, enhancing the efficiency by two orders of magnitude. Besides, this model yielded NegJ similar to that of the B-Spline scheme. This suggests that the DVF extracted using this model had comparable physical plausibility with that obtained using the B-Spline scheme. Conclusions The Transformer-based deep learning model for automatic target volume tracing fills the functional gap of the Elekta Unity MRI-guided radiotherapy system, facilitating relatively accurate, efficient automatic tracing of moving tumor targets in the chest and abdomen. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|