An Algorithm For Locating Bone And Joint Anatomical Points In CT Images Based On Deep Learning

The location of anatomical signs is an important step in the analysis of medical images,and a pre-step in many medical analysis methods.At the same time,it can help doctors to carry out clinical diagnosis and provide convenience for doctors to further judge the condition.An algorithm for automatic location of hip and knee joint anatomical points in CT based on deep learning is proposed.The positioning algorithm is divided into four steps.The first step is preprocessing,which mainly truncates and normalizes the image,removes the metal background of the image,and samples the image to the same specification through isometric resampling,so as to provide better data for deep learning.The second step,rough ROI positioning of anatomical points,is aimed at the problem of too large CT data.After reducing the image resolution,traditional image algorithm is used for morphological analysis.The 3D image is mapped into binary images of three planes.After removing small connected domains and numerical filtering,the coordinate points of each part are calculated to get the center of ROI region.Then,multiple high-resolution local images are obtained by clipping with the center points as the center.The third step uses local image as deep neural network input to train the fine location of anatomical structure.The fine location network is based on the SCN network,and the following improvements are made: first,the accuracy loss and data offset of traditional heat map regression are eliminated by introducing the differentiable space numerical transformation.The differentiable space numerical transformation is an operation method to calculate coordinates from heat map without increasing the complexity of network.Secondly,on the basis of U-shaped network,the heat map regression convolution kernel is increased to enlarge the image receptive field.Finally,the fourth step carries out postprocessing on the prediction results,converts the network prediction results into the position output of the original image before processing and calculates various evaluation indexes.A total of 28 sets of normal data and pathological data were selected from the threedimensional anatomical CT images of hip and knee provided by United Pictures for the experiment.The experimental results show that the accuracy and stability of location can be achieved.