Study On The Algorithms For Image Feature Extraction In Similar Pulmonary Nodules Search

In the course of clinical diagnosis of lung disease in the presence of a pulmonary nodule,doctors usually refer to the diagnosis of similar pulmonary nodule images.Research on content-based image search method and the realization of similarity search system of lung nodules will help to provide reference to doctors to improve the diagnostic efficiency and accuracy.A core work in content-based image search systems is the feature extraction of pulmonary nodule images.We have implemented and compared three feature extraction schemes.First,based on the previous work on the detection of pulmonary nodules in the laboratory,a series of low-level and high-level features were extracted by using the method of manual design.The two is the use of deep convolution neural network.The images of two nodules are spliced together to form an image,and their similarity values are taken as the categories.The image classification method is used to train the network,and then features were extracted from the middle layer of the network.Three is the use of Twin Towers(Siamese)convolutional neural network structure.The two towers have the same network nodes and share the weights of the network.Each of them is processed by the input image.The network is trained by back-propagation of the loss function value,and the output of the middle layer of the network is extracted as the image feature after the training.The overall layout of the program is analyzed,which is divided into 5 major components,the storage structure blob,layer Layer,network net,network optimization scheme slover and the proto format for the structure,storage and reading of the network model.Net forms a directed acyclic graph(Referred to as DAG)in combination of Layer and slover implements a method for solving DAG.Based on the Caffe framework,you can add the implementation function of the custom layer.In order to better understand convolution kernels,we reconstruct the image by using the convolution kernel,and verify the validity of the convolution kernel.The data used in this project include the LIDC-IDRI(Lung Image Database Consortium and Image Database Resource Initiative)database and the scores of similar pulmonary nodules pairs which were provided by the cooperative unit.The three methods are used to extract the features of the similar image retrieval.The results show that the features extracted by Siamese are the best.