The Research On Content-based Medical Image Retrieval Algorithm

CBIR plays an important role in clinic, teaching, research and PACS (Picture Archiving and Communication Systems), etc. the dissertation centers on crucial technologys of content-based medical image retrieval technology.The dissertation studies architecture and the key technologies of content-based medical image retrieval system which includes image shape and texture features extracting and expressing, image matching and relevance feedback, performance evaluation and so on.Medical images supply us with much texture information, so texture plays an important role in image recognition. Fractal feature is introduced to extract feature of medical image. This method discovers well at fractal dimensions of the normal lung and several kinds of common lung diseases. In order to get all the texture information of each scale and orientation, the dissertation adopts the method of gabor multi-resolution wavelet analysis transform, then, the dissertation adopts improved image's fractal dimension to calculate the images after filting. In order to increase the texture information, the dissertation takes account of Tamura fatures. The experimental results indicate that the method has high recall, precision and average sorting.The dissertation proposes image shape-retrieval method based on the wavelet modulus maxima and combined morphology, in the case of overcoming the ignored low-frequency information and the problem of discontinuous boundary. Method of wavelet modulus maxima is applied to high-frequency part of wavelet decomposition image. Low-frequency part is based on morphology. Last, the two boundaries of image are fused. The experimental results indicate that the boundary of the image is clear, continuous and smooth.In the same time, the dissertation proposes a new method for image retrieval using combined texture and shape features. The algorithm of relevance feedback adopts the method of adjusting the intra-weight, inter-weight of the texture and shap features. According to feedback information of users', if an image contains more texture information, it gets more weight of texture features; and it is the same to shape features. The dissertation also adopts graded retrieval method which reduces computational complexity on the precondition of the accuracy. Experiments show that the results obtained from combined-features are better than the retrieval results obtained from single-features.