Research On 3D Point Set Registration Algorithm For Lung CT Images

Image-guided radiotherapy is an important method for the treatment of lung cancer.In the process of image-guided radiotherapy,the interference of respiratory motion and heart beat is an important factor that causes image artifacts.Accurate lung tissue motion estimation can reduce the artifacts caused by respiratory motion and improve the accuracy of radiotherapy.Image registration is a key technique for lung motion estimation.In order to solve the problem of insignificant feature extraction and inaccurate anatomical structure in the point set matching,which affects the registration accuracy and registration speed,this paper studies the feature extraction algorithm and the three-dimensional points of lung CT images Set registration algorithm,the specific research content is as follows:(1)Aiming at the significant anatomical structure of the lung trachea bifurcation points and the number remaining unchanged,the displacement vector can fully reflect the spatial transformation in image registration.A method for feature extraction of lung trachea bifurcation points was proposed.The lung trachea is extracted and segmented,and the geodesic distance from each pixel to the tracheal wall is calculated,and the number of vessels(3 or more)connected to the lung trachea to detect the lung trachea is set as the connection point.Finally,the connection farthest from the tracheal wall is selected by clustering The point serves as a tracheal bifurcation point.(2)In order to solve the problem that the point set matching processing a large amount of data leads to a long running time.This paper proposes a deep learning-based coherent point shift(CPD-Net)algorithm for unsupervised learning of geometric transformations.Using multi-layer perceptron(MLP)and activation function,fitting a continuous and smooth displacement function,and learning the displacement function to estimate the geometric transformation through the training set.In the non-iterative optimization process,the data in various dimensions can be improved.Computing performance.In the lung CT image(DIR-Lab data set),the labeled points or extracted feature points are used as the data set.The simulation comparison between the CPD-Net algorithm and the coherent point drift algorithm shows that the CPD-Net has better robustness and Accuracy,registration speed within 1 second,with good real-time.