Research On Medical Image Retrieval And Compression

Medical imaging is an important basis to assist clinicians in diagnosis,treatment and analysis of diseases.With the rapid development of digital hospitals and the emergence of more and more advanced medical imaging equipment and technology,a large number of medical images are produced in hospitals every day,which increases greatly the storage space of medical image data,and also the workload of radiologists.In addition,the diagnoses made by radiologists are often subjective,and thus prone to misdiagnoses and missed diagnoses.Retrieval and compression of a large amount of medical image data can reduce the workload of diagnostic physicians,improve the diagnostic efficiency and accuracy of diagnostic physicians,and reduce the demand of medical image data on storage space.Therefore,this thesis studies the methods of medical image retrieval and compression,and the research contents are as follows:(1)Medical image hash retrieval combined with dense convolutional neural network.Image feature extraction and hash encoding are two key steps in medical image retrieval.In order to extract image feature that can fully express the image content and reduce the projection error in hash encoding,a method based on dense convolutional neural network(Dense Net)and improved supervised kernel hash is proposed.This method firstly optimizes the Dense Net model and extracts the high-level semantic features of the images with the trained Dense Net model.Then,the kernel principal component analysis projection is performed on extracted image features,and the nonlinear information in the image features is fully exploited to reduce the projecion error.Then supervised kernel hash learning is carried out on the projected image features,which are mapped to hamming space to generate a more compact binary hash code,and the retrieval is completed by comparing the hamming distance between the hash codes.Experimental results show that,in the LUNA16 data sets and Paris6 K data sets,compared with the other six common hash methods,the average retrieval accuracy values of this paper method are higher than other methods under different hash code length,when the hash code length is 64 bits,the average retrieval accuracy value reaches a maximum of 92.9% and 89.2% respectively,so as to verify the validity andexpansibility of this paper method.The time complexity of this paper method is compared with that of the hash algorithm based on convolutional neural network,the result shows that the time complexity of this paper method is smaller,which indicates that this paper method has certain efficiency.(2)A medical image compression method combined with Canny edge detection and SPIHT.In view of the particularity of medical image,high-frequency information such as texture details in images is crucial for diagnostic physicians to make judgments.Therefore,in this paper,starting from improving the retention of high-frequency information in the recinstructed image after compression,and aiming at the deficiency that the reconstructed image of Set partitioning in Hierarchical Trees(SPIHT)algorithm will lose high-frequency information,a medical image compression with Canny edge detection and SPIHT is proposed.In this paper method,Canny edge detection is carried out firstly to extract the high-frequency information of the image and obtain the edge image of image.Then,the image is encoded by SPIHT,and the stream encoded by SPIHT is optimized with Huffman algorithm.After Huffman decoding and inverse wavelet transformation,a reconstructed image is obtained,and the reconstructed image is combined with the edge image obtained before to restore the the original image.Experimental results show that,in the selected 10 includes medical and non-medical standard test images,compared with the algorithm of combining SPIHT and Huffman,the information entropy of the reconstructed image of this paper method is higer under different bit rate.And it is shown that the high-frequency information of reconstructed image of this paper method is effectively retained,which verifies the feasibility and effectiveness of this paper method.