Detection Of Bifurcation Points Of Pulmonary Tracheal Trees In CT Images Based On Deep Learning

The classification of pulmonary tracheal tree in human lung CT images is very important for the study of assisting the diagnosis of lung diseases.The detection of key bifurcation points of pulmonary tracheal tree is the key step in the classification of tracheobronchial tree.Using deep learning method to directly detect the key bifurcation points of the pulmonary tracheal tree has research significance for the classification of the pulmonary tracheobronchial tree,and has very important practical significance for the subsequent auxiliary diagnosis of pulmonary diseases.In view of the different difficulty of detecting different bifurcation points of pulmonary tracheal tree in human lung CT images,a parallel residual network which is named Multiple-ResNet for detecting key bifurcation points of pulmonary tracheal tree is designed on the basis of ResNet,which combines the three-dimensional structural characteristics of pulmonary tracheal tree.Multiple-ResNet combines three subnetworks which include F,L2 and R3.Network F extracts global features to detect four key bifurcation points by using modified three-dimensional ResNet50 structure;Network L2 and R3 both extract local features to detect local key bifurcation points by using modified three-dimensional ResNet34 structure.The prediction results of network L2 and R3 revise the prediction results of network F to improve the detection accuracy of four bifurcation points.Four hundred pulmonary CT images in LUNA 16 data set are manually labeled with bifurcation points of the pulmonary tracheal tree in order to obtain the data set of bifurcation points of the pulmonary tracheal tree.Considering the different location and size of the tracheobronchial tree in different lung CT images,a data preprocessing method for detecting the bifurcation points of the pulmonary tracheal tree in human lung CT images is proposed.First,according to the lung mask extracted from the original lung CT image,the lung bounding box is intercepted from the original image after the reciprocal processing as the region of interest to reduce the impact of the useless information of the lung.Then the input data of three sub-networks are generated.Finally the input data are normalized to improve the convergence speed and performance of the network model.Considering the characteristics of the bifurcation points of the pulmonary tracheal tree,the hyper-parameters suitable for learning the bifurcation points data characteristics of the network model are set up in order to improve the training effect of the network model.Seventy human lung CT images were randomly selected to test the deep learning method which is proposed,and the results were compared with those of traditional algorithms and deep learning methods with different network structures.The experimental results show that the deep learning method based on Multiple-ResNet network model can detect the four key bifurcation points of pulmonary-tracheal tree.