Data Clustering And Medical Image Analysis Based On Subspace Analysis

Over the past few decades,data in multiple modalities has exploded.Although these data are mostly high-dimensional in form,their intrinsic dimensions are usually much smaller than the formal dimensions,and subspace analysis,which is widely used at present,can be used to analysis and process the intrinsic dimensions of high-dimensional data.In subspace analysis,the inter-cluster correlation which is caused by noise will affect the analysis of the internal structure of the data,resulting in the performance degradation of subspace algorithm.Therefore,this paper improves the subspace analysis algorithm of low rank representation model based on the correlation between different subspaces.Meanwhile most of the current breast classification algorithms are based on single level classification algorithm,so this paper proposes a two-level classification algorithm and applies the subspace method to the analysis of medical image data.The work of this paper is mainly divided into the following two parts.(1)A method for detecting the inter-cluster correlation is proposed,and cluster analysis is performed on this basis to to reduce the influence of noise such as illumination on subspace analysis.This paper detects abnormal data with strong inter-cluster correlation based on the representation matrix obtained from the low-rank representation,and constructs a new dictionary to establish a low-rank model and perform subspace clustering on the data.In addition,it is proposed to add preprocessing to texture data,which allows subspace clustering to be effectively extended to texture data analysis.Experimental results show that the clustering accuracy of the algorithm on texture data is 90%,the clustering accuracy on Extended Yale Database B is 93%,and the clustering accuracy in the two experiments on the AR dataset are 95% And 83% respectively.At the same time,in order to prove that the proposed method can reduce the inter-cluster correlation,the correlation parameters on different datasets are calculated,and it is found that the correlation parameters in the proposed method are the smallest.(2)In the classification of benign and malignant breast DCE-MRI tumors,based on morphological features,dynamic enhancement features and texture features of the lesion,a two-level classification model is proposed to improve the overall classification performance by distinguishing common lesions and hard-to-spot lesions.Subspace analysis model is further used and the dataset is divided to obtain a reliable sample set of hard-to-spot lesions,then reliable classification results can be obtained.First,a two-level classification model for breast benign and malignant classification is proposed.The classification results of the three types of features in the first-level classification are used to find hard-to-spot lesions and common lesions are classified.Then Then,feature selection and fusion of various characteristics of the lesions are carried out to classifying the hard-to-spot lesions twice.The results are that sensitivity is 94%,specificity is 90%,accuracy is 92%,and Matthews correlation coefficient is 84%.In the classification model,the selection of hard-to-spot lesions is an important factor affecting the final classification performance.Therefore,this paper use subspace analysis to obtain the sample set of hard-to-spot lesion.This method clusters the data by using a low-rank representation model to obtain data with abnormal distribution,that is,hard-to-spot lesions,and then the common and hard-to-spot lesions are classified respectively.Finally,the results of the classification are that sensitivity is 100%,specificity is 93%,accuracy is 97%,and Matthews correlation coefficient is 94%.The experimental results show that the subspace analysis method to find hard-to-spot lesions can further improve the classification performance of breast lesions.The research in this paper shows that the clustering performance can be improved by reducing the inter-cluster correlation,and this paper also obtains ideal results based on subspace analysis in medical image.