Research On Content-based Retrieval Of3D Medical Images

With the development of modern medical imaging technology and computertechnology, medical images are being used more and more widely. Large numbers ofmedical images are produced in hospitals every year. Therefore, efficient methods oforganizing and retrieving these images are needed to track disease progression, search forsimilar cases and aid medical diagnosis, therapy and research.Content-based image retrieval (CBIR) has been a vivid research field in recent years.Some real applications have been put into use. As for medical images, there are alsomany early researches, but most of these researches focus on2D images.3D medicalimages, such as CT and MRI images, are taking a large portion in medical application.However, there are few researches on3D medical image retrieval. This paper focuses on3D medical image retrieval and studies key techniques in feature extraction andcomparison, from both aspects of local structure classification and image similarity.Experiments are performed on real images collected from hospitals, with performanceevaluation of experimental results.For the retrieval method based on local structure classification, an approach isproposed to classify local structures into line-like, blob-like and sheet-like patterns indifferent sizes. This approach is based on the second-order derivative in multi-scale spaceusing eigen-analysis of Hessian matrix. Properties of eigenvalues on different patternsare analyzed and normalized response functions are proposed. Tissue structures in imagesare classified into three categories according to the outputs of response functions.Extraction and composition methods in multi-scale space are also studied. Retrieval ofspecific structures in3D images by pattern and size is implemented.For the retrieval method based on image similarity, considering the characteristics ofmedical images, a new method combining visual features is proposed to retrieve2Dslices of3D images. This method uses the combination of intensity feature by cumulativehistogram, texture feature by Gabor transform, edge feature by Canny detector and localstructure feature by second-derivative method. Image similarity is computed bycalculating the distance of combined feature vectors. Retrieval of similar images fromimage library by a given image example is implemented.