Research On Content-Based Medical Image Retrieval Of The Overall Similarity

Along with the rapid development of digital imaging equipment, such as CT, MRI, X-ray, US, PET and so on, today’s medical institutions produce enormous amounts of digital medical images every day. The amount of digital images in large medical institutions grows exponentially. Thus the needs of high efficient storage and management tools are increasingly urgent. As traditional text-based retrieval cannot fully meet the applications, content-based image retrieval techniques are getting more and more attention from researchers, and become one of the most vivid research areas.Based on analyzing key techniques and development trends of content-based medical image retrieval systematically, this dissertation focuses on the overall similarity retrieval of medical images, and gets good performance. The main works of this dissertation include:(1) To retrieve the overall similar images from medical images which comprise several body parts. Rigid registration based retrieval method can get very high accuracy, but is difficult for application, because of its high time consuming. For this problem, this dissertation proposes a method that uses gradient phase mutual information and the image gravity center to estimate the angle deviation and the translation deviation. Experimental results show that the new algorithm can provide the same precision as rigid registration while reducing retrieval time dramatically.(2) To retrieve same anatomical regions from medical images which comprise single body part. This thesis proposes linear weighting method which integrates with priori knowledge about the images being retrieved. The method is based on Powell optimization algorithm, and arranges similarities reasonably according to the priori. The method trains Powell with a small amount of images, optimizes different weights of different similarities, and gets a combined similarity measure. The combined similarity measure achieves good accuracy, and the optimization process of Powell converges quickly. (3) To retrieve same annotation class from medical images which comprise different body parts. This dissertation combines many global and local features’ similarities to improve retrieval performance. First, the codebook is made by separately clustering. This method shortens computation time significantly while maintains good accuracy. Then, the "sum rule" and the "product rule" are compared on public datasets by combining large amount of similarities. Experiment results show that the accuracy of the "sum rule" is better than the "product rule" on average, and theoretical analysis is also taken. The conclusion provides reliable experience for real applications.(4) To improve the retrieve results of same annotation class from medical images which comprise different body parts. This dissertation introduces rank-SVM to promote the accuracy of the retrieval result. By filtering out the redundant similarities using mutual information, the method takes the remains as the input of rank-SVM to improve the rank result. Experiments and analysis show that the method improves the retrieval accuracy for medical images significantly, and is proper to application with its relatively lower computation consuming.