METHODS: The method models the 3D motion of each voxel as a linear displacement along a direction vector, with a variable amplitude and phase that depend on the voxel location. The direction vector and amplitude are derived from the registration of the CT images at the end-of-exhale (EOE) and the end-of-inhale (EOI) phases. The voxel phase is estimated by a neural network. Coordinate convolution (CoordConv) is used to fuse multimodal data and embed absolute position information. The network takes the front and side views as well as the previous phase views as inputs to enhance accuracy.
RESULTS: We evaluate the proposed method on two datasets: DIR-Lab and 4D-Lung, and obtain average errors of 2.11 mm and 1.36 mm, respectively. The method achieves real-time performance of less than 7 ms per frame on a NVIDIA GeForce 2080Ti GPU.
CONCLUSIONS: Compared with previous methods, our method achieves comparable or even better accuracy with less CT phases.
方法:该方法将每个体素的3D运动建模为沿方向矢量的线性位移,具有取决于体素位置的可变振幅和相位。方向矢量和幅度是从呼气末(EOE)和吸气末(EOI)阶段的CT图像的配准中得出的。体素相位由神经网络估计。坐标卷积(CoordConv)用于融合多模态数据并嵌入绝对位置信息。网络采用前视图和侧视图以及先前的相位视图作为输入以提高准确性。
结果:我们在两个数据集上评估了提出的方法:DIR-Lab和4D-Lung,并获得2.11mm和1.36mm的平均误差,分别。该方法在NVIDIAGeForce2080TiGPU上实现每帧小于7ms的实时性能。
结论:与以前的方法相比,我们的方法以更少的CT相位实现了相当甚至更好的精度。