Based On The Content Of Multi-modal Medical Image Retrieval System

With the rapid development of the Internet and the popularization of various digital equipment, image data growths explosively, thus management and retrieval of them becomes increasingly difficult. The emerging Content-Based Image Retrieval (CBIR) technique is born to solve the problem. A CBIR system constructs image identities based on contents of images such as color, texture and shape by which retrieval similar images to the query image in an image database, and shows the result to users.Medical image is very important for clinical diagnosis process. Picture Archiving and Communication Systems (PACS), which based on modern digital imaging technology and Digital Imaging and Communications in Medicine (DICOM) protocol, is a technique that integrates the obtaining, diagnosing and achieving of digital medical images. Until now, retrieval in the network of DICOM is according to text information. It is apparently that indigent text information cannot substitute images with abundance information. Moreover, conventional text retrieval has restricts the application of medical image greatly. Therefore, research of CBIR technique will bring great revolution into the exploitation of medical image database.We analysis the characters of multi-mode database deeply, try to improve the existing PACS. We present a layered retrieval framework, firstly to classify the query image, then to retrieve in class. Under such a framework, the time spend on retrieval do not increase along with the scale of database, it relates to the number of images in each class. This framework improves ability of retrieval system working in real time greatly. For retrieval in class, we can design adaptive scheme, which can bring great efficiency to the whole system.There are some kinds of noise appearing in medical image frequently. Such as, ①White margins leave behind when the negatives are turned to digital images ② Images slant ③ There are texts and logos in the image. According to the characters of these noises, we design a series of method to denoise. We use Hough transform to detect the edge of white margin, and then skew correction, at last text and logo removal.For the classification, we reduce dimension of features firstly, then classify by near neighbor classification. In order to reserve enough information of space distribution and classify best, we choose the Fisher criterion to reduce dimension. Considering data distribution is non-linear, we choose GDA method which using kernel technique. At last, we compare GDA and LDA method, the result indicates that GDA method is more effective.Retrieval is the most important matter in CBIR system, in this aspect, we use Gabor feature to retrieve "chest" X radiology images. Gabor filter is a powerful tool in vision feature extraction, Gabor filter is acknowledged as the best method for texture feature. In order to reducing compute, we sample the image after convolution adaptively. We partition the whole image into blocks, then use PCA to reduce dimension. At last, we use Precision rate and Recall rate to evaluate the retrieval system.