Research On Medical Point Cloud Super-resolution Algorithm Based On Graph Convolutional Networks

Point clouds are a widely used form of 3D data.Medical 3D models based on point clouds have important medical reference value in various applications such as disease detection,treatment planning,and surgical navigation.In the actual process of acquiring medical point clouds,the limitations of device resolution,the complex three-dimensional structure formed by the mutual occlusion of target tissues,and the computational accuracy of reconstruction algorithms will affect the data quality of point clouds.Point clouds obtained from scans or reconstruction calculations are often sparse and non-uniform,with noise.Aiming at the problem of insufficient resolution of medical point clouds,this study designs a point cloud super-resolution algorithm based on deep graph convolution networks.This study constructs two duarl-modality medical datasets containing both high-resolution 3D models and low-resolution data.The contents and results are as follows:（1）Aiming at the shortcomings of insufficient local feature extraction ability and poor multi-scale feature fusion ability in point cloud super-resolution network,this paper designs an upsampling module based on Dense GCN point cloud feature extraction module and Non Local multi-scale feature fusion.Datasets and high-resolution 3D models of tooth datasets are used for experiments,and the effectiveness of the network improvement in this paper is verified by multiple test data.The network proposed in this paper achieves point2surface distance,Chamfer Distance and Hausdorff Distance on dense intraoral scanning point clouds of 0.0583,0.0289*10^-3,0.6158*10^-3respectively.point2surface distance,Chamfer Distance and Hausdorff Distance on dense blood vessel point cloud They reached 1.1394,0.2301*10^-3,10.2983*10^-3respectively.（2）In this paper,a dual-modality blood vessel dataset of two-dimensional blood vessel images and high-precision blood vessel three-dimensional models is constructed.A reconstruction method of super-resolution 3D model of cerebral blood vessels based on 2D vascular images was designed.The reconstruction method includes a sparse point cloud reconstruction method based on two-dimensional blood vessel images and a depth map convolution point cloud super-resolution network model trained on a high-precision three-dimensional model of blood vessels.The sparse point cloud reconstruction method includes multi-scale fusion filtering vessel segmentation algorithm and three-dimensional active contour model reconstruction vessel point cloud algorithm.The experimental results show that the reconstruction method designed in this paper can effectively realize the three-dimensional reconstruction of two-dimensional blood vessel images,and provide a new idea for the three-dimensional reconstruction of multi-plane two-dimensional angiography images.（3）In this paper,a CBCT and intraoral scan bimodal tooth 3D model dataset was constructed.A super-resolution method of CBCT tooth 3D model was designed,including registration of CBCT tooth 3D model and intraoral scanned tooth 3D model and training depth map convolution point cloud super-resolution network model.The experimental results show that the CBCT super-resolution method designed in this paper can effectively improve the resolution of the CBCT tooth 3D model and meet the requirements for the resolution of the 3D tooth model in oral medicine.