Medical Image Retrieval Based On Low Level Features And Semantic Features

With the development of science and technology, magnetic resonance imaging, CT and other medical imaging devices have made a major breakthrough, for example, 3.0T magnetic resonance imaging equipment has become a mainstream device. Medical images has become an important tool for doctors to determine the condition and diagnosis, hospital stores a lot of patient medical imaging data, forming a large capacity medical image database. It becomes a serious problem to how to help doctors retrieve the target medical images more quickly and accurately.Content-based image retrieval, namely CBIR, is the mainstream of development in the field of computer vision about image retrieval. CBIR system works like this: user input one sample image, then CBIR system calculates and outputs target images whose contents are the same or similar with the input image. In this article, we focus on content-based medical image retrieval namely CBMIR. Using CBMIR to retrieve similar images and obtain diagnostic data of the patients, CBMIR can provide treatment options for doctors.CBIR including two key technologies, image feature extraction and image feature matching. Image feature extraction is divided into two categories: low-level features, its contents include color, shape and texture; A priori knowledge to analyze image information to obtain image semantic content. Calculating distances between two vectors extracted from two images, the smaller the distance, the more similar the two images.The process of CBMIR is, first, using the global gray hash algorithm to extract gray features from the medical images, using Gabor wavelet algorithm to extract texture features from the medical images, using canny operator to extract shape features from the medical images. Then adjusting the weights of the images feature vectors used by genetic algorithm, and retrieving similar medical images based on multi features of medical image features. In order to reduce the semantic gap between pathological similarity understood by doctors and visual similarity understood by computer, we need to adjust the feature weights again using genetic algorithm based on feedback from the doctor. Finally, we successfully retrieve similar medical images.