Research On Medical Image Retrieval Algorithm Based On Geometrical Relationship Of Features

Medical images have always been one of the important clinical bases for doctors in diagnosis and treatment.With the continuous progress of medical imaging technology and the upgrading of acquisition equipment,its importance is increasingly prominent.How to analyze and utilize the massive medical images effectively becomes a key problem.The introduction of image retrieval technology is an effective method to solve this problem.Text-based image retrieval method and content-based image retrieval method are widely used in medical image retrieval.Text-based medical image retrieval method,which describes medical images by manual annotation,has more subjective factors,which limits the retrieval effect to a certain extent.Content-based medical image retrieval method extracts the content information of images by computer automatically and describe medical images using the extracted medical image features.This method avoids the influence of subjective factors and the extracted information is rich,so the content-based medical image retrieval method becomes a research hotspot.At present,the medical image features widely used in medical image retrieval based on content include shape,texture,pixel strength and SIFT.It is the key link of this method to analyze the features of medical images and generate the description of image features.Among them,BoV model,Visual Vocabulary Tree model and VLAD model are commonly used.In the feature extraction and similarity measurement of medical image,the geometric relationship between feature points is often ignored,which leads to the occurrence of mismatching and affects the retrieval effect.Aiming at this problem,a medical image retrieval method based on geometrical relationship of features is proposed in this paper.On the basis of the medical image retrieval method based on vocabulary tree,the geometric verification link is added.By verifying whether the matching feature points conform to the geometrical relationship rationality to reduce the occurrence of mismatching,improve the retrieval accuracy.We also proposed a spatial pixel strength feature extraction method based on geometric constraints and combine this method with the VLAD model.By extracting the pixel intensity information of each feature point and the related feature points in its specific neighborhood,the method can quantify and describe the feature point and then generate the VLAD model after cluster analysis.Comparing with extacting each feature point's pixel strength information separately,this method of extraction can describe the feature points more sufficiently.The advantages of high efficiency and high accuracy of VLAD model have further improved the retrieval performance of this method.We verified the retrieval performance of the two methods in the DDSM breast image database and human brain tumor MRI database.Compared with the method of breast cancer image retrieval based on vocabulary tree,we added the geometric verification algorithm after we got the initial results to reduce the error matching and improve the retrieval accuracy.After adjusting the parameter value setting of the algorithm,we found the appropriate parameter values.In the experiment of brain tumor image retrieval based on geometric constrainted spatial pixel intensity,we compared our method with the method based on BoV model with pixel intensity,the method based on BoW model with SIFT and the method based on VLAD model with SIFT.The experimental results show that our method have better performance.After comparing the different parameter values of the algorithm in experiment,we found the appropriate parameter values.