The Research Of Automatic Feature Extraction And Fuzzy Feature-based Image Retrieval Of Medical Images

With the wide use of digital medical devises such as CT,MR,X-Ray and so on, hospitals produce a great number of digital images everyday.PACS(Picture Archiving and Communication Systems) are used in more and more hospitals to store and access the ever-increasing images.Diagnosis for some unfamiliar diseases, pathology researches and medical education require searching similar images from database and this is a very hard and boring work and usually needs the help of artificial intelligence.In most of PACS,the images are organized with text,allowing the images to be accessed by text-based searching.However,with the emergence of massive image databases,the traditional text-based search suffers from dependence on subjective and manual annotations.To overcome these difficulties,the techniques of content-based image retrieval(CBIR) have been major topics of research for medical image database queries instead of text-based searching techniques in recent years.Significant developments of CBIR techniques have been made in research and commercial applications since 1990's.However,due to the low resolution and strong noise of medical images represented in gray level rather than color,content-based retrieval for medical images is still faced with challenges.In the process of radiology medico-diagnosis,clinic determinations are usually based on the regional anatomic and physiologic information in medial images(Region Of Interest,ROI).In order to extract the regional features automatically,each of the medical images in database is segmented. To overcome the uncertainty of segmentation,the fuzzy features are used in some CBIRs.This paper focuses on the researches of automatic brain ROI features extraction of serial brain images and retrieval of brain images with fuzzy region features.In the research of automatic brain image feature extraction,a region growing and morphology based approach is developed to get brains from series of cerebral Computerized Tomography with the knowledge of anatomy.The seed pixels are selected along a helix the starting point of which is the centroid of the image when growing.A modified BET(Brain Extraction Tool) algorithm is also proposed to extract brains from series of MR images in this paper.The modified algorithm simplifies the smoothing force used in BET which makes the contour of edge smooth and modifies the expansionary force used in BET to evolve the edge of brain according to the intensity distribution and the gradient of images.The modified expansionary force puts the contour fast when the contour is in the brain and puts the contour slowly when the contour is close to the edge of brain,which can resolve the edge overflow problem in BET.To overcome the problem that the parameters are unstably renewed,this paper also proposes a parameter-limited GMM-EM algorithm to segment the brain images fast and stably.In the research of fuzzy feature based retrieval algoritm,this paper totally extracts the average intensity features,wavelet texture features,Gabor texture features and moment shape features of each region,and convert these crisp features into fuzzy features with exponential membership function.Compared with other membership functions,the exponential membership function is more easily calculated and extended.Based on the exponential membership function,the measurement of fuzzy similarity between two fuzzy features is derived.The similarity between two region represented by multi-fuzzy-features is also defined based on the exponential membership function.To reduce the influence of segmentation uncertainty,this paper presents a new method to retrieve cerebral CT images based on fuzzy binary tree structure.This method first segments the images into several regions,then gets the fuzzy binary tree structures by assigning each region a membership degree which a pixel belongs to the region according to the intensity standard deviation of each region and gets the fuzzy intensity features, fuzzy wavelet texture features,fuzzy Gabor texture features and fuzzy moment shape features of each region by exponential membership function.Based on the classical fuzzy region-based image retrieval algorithm UFM,the regions in two images are matched and the similarity between two images is calculated by these fuzzy content features and fuzzy binary tree structure features.Compared with UFM,our method uses not only the fuzzy content features but also the fuzzy structure features,and our method can merge the regions according some rules when matching the regions and needn't calculate the features of the merged region again.Experiments showed that this method is robust to the uncertainty of image segmentation.To promote the retrieval precision furthermore,this paper integrates the classical re-weighing relevance feedback algorithm into the UFM algorithm and proposes a new fuzzy region-based relevance feedback algorithm.This algorithm uses the relevant images and maximizes the weighted product of fuzzy region similarity between the query image and relevant images to get a weighting vector which assigns weights to the features representing each region and symmetric matrixes to transform the features into new optimum spaces which the user wants.This paper also combines the SVM based relevance feedback algorithm using the global wavelet energy features of both the relevant images and irrelevant images with the proposed fuzzy region-based relevance feedback algorithm.Experiment shows that the hybrid algorithm performs better.