r-UNet: Leaf Position Reconstruction in Upstream Radiotherapy Verification
r-UNet: Leaf Position Reconstruction in Upstream Radiotherapy Verification
复制标题
DOI：
10.1109/trpms.2020.2994648
复制
发表时间：
2021-03-01
影响因子：
4.4
通讯作者：
Hugtenburg, R. P.
中科院分区：
文献类型：
--
作者：
De Sio, C.;Velthuis, J. J.;Hugtenburg, R. P.
单片有源像素传感器（MAPS）设备是一种有效的工具，用于调强放疗（IMRT）治疗的上游验证。为了提高治疗的质量和安全性，真实的实时高精度地测量用于对射束进行整形的多叶准直器（MLC）的位置是至关重要的。本文描述了r-UNet，一种基于深度学习的叶位置重建解决方案。该模型用于分析Lassena MAPS设备产生的高分辨率图像，以自动确定叶片位置。图像分割和叶片位置估计是在多任务环境中同时进行的。r-UNet获得了保留测试集中重建图像掩模的平均Dice系数为0.96 +/-0.03;而由MLC位置的估计产生的均方误差（MSE）为0.003mm，对于1和35 mm之间的叶片延伸，分辨率范围在45和53 μ m之间。在看不见的叶片位置上，r-UNet产生了一个单一的叶分辨率在54和88 μ m之间，这取决于叶的延伸，平均MSE为0.07 mm。这些结果是使用以34帧/秒收集的单帧数据获得的。
Monolithic active pixel sensor (MAPS) devices are an effective tool for upstream verification of intensity-modulated radiotherapy (IMRT) treatments. It is crucial to measure with high precision the positions of the multi-leaf collimators (MLCs) used to shape the beam in real time, in order to enhance the quality and safety of treatments. This article describes r-UNet, a deep learning-based solution for leaf position reconstruction. The model is used to analyze the high-resolution images produced by a Lassena MAPS device in order to automatically determine the leaf positions. Image segmentation and leaf position estimation are performed simultaneously in a multitask setting. r-UNet obtained an average Dice coefficient of 0.96 +/- 0.03 for the reconstructed image masks in the held-out test set; whilst the mean squared error (MSE) resulting from the estimation of the MLC positions is 0.003 mm, with a resolution ranging between 45 and 53 mu m for leaf extensions between 1 and 35 mm. On unseen leaf positions, r-UNet yielded a single-leaf resolution between 54 and 88 mu m depending on the leaf extension, and an average MSE of 0.07 mm. These results were obtained using single frames of data collected at 34 frames/s.