Algorithms Of CT Image Retrieval With Multi-feature Representation

Computed Tomography examination is an important means of disease diagnosis.In order to make a stable and reliable disease diagnosis for the current patients,doctors often need to refer to the CT images of historical medical records.However,with the huge and complex CT image databases,how to retrieve the CT images required by doctors with high accuracy has become an urgent problem to be solved in the current computer-aided diagnosis technology.Due to the different shape of the CT image,the results of traditional CT image retrieval based on single low-level visual feature are not satisfactory.Therefore,in order to improve the accuracy of CT image retrieval results,this thesis studies the feature representation of CT images from the perspective of multiple features,and proposes two CT image retrieval algorithms based on multiple feature representations.1)Due to traditional feature fusion methods are difficult to effectively fuse homogeneous local features and heterogeneous local features of CT images,this thesis proposes a CT image retrieval algorithm based on fine-grained correlation analysis.This algorithm first performs CT image segmentation processing to obtain fine-grained samples that contain different local features.Then,we use the K-means clustering methods to label fine-grained samples that have similar characteristics with the same label.We build the correlation function between different low-level visual features of fine-grained samples,and obtain the fusion features of fine-grained samples through correlation function optimization processing and feature fusion strategy.Finally,the fusion features of all fine-grained samples from the CT image are recombined together to construct the final CT image feature representation.The experimental results show that the average retrieval precision on the datasets EXACT09,TCIA,and NEMA are: 0.9312,0.9717,0.9911,respectively.2)Due to traditional image hashing methods are difficult to calculate image data with ultra-high dimensions and cannot combine multiple features for image hash coding,this thesis proposes a CT image retrieval algorithm based on joint sub-space hashing.This algorithm first performs segmentation processing on the CT image to obtain a series of local blocks.Then,we extract three kinds of subspace projection matrices for local blocks:principal component subspace projection matrix,two-dimensional principal component subspace projection matrix and local reserved subspace projection matrix.Then,the three subspace projection matrices are combined to construct the objective function of the joint subspace hash algorithm based on the joint sparse hash method.Finally,the objective function is optimized through the alternating optimization algorithm,and the hash representation of the CT image is obtained.The experimental results show that the average retrieval precision on the datasets EXACT09,TCIA,and NEMA are: 0.9347,0.9735,0.9924,respectively.