Research On Multi-Modality Image Retrieval Of Brain Diseases

With the development of the medical imaging technology, and as the diseases representations in images are a significant reference to diagnosis, medical images are playing a more and more important role in diagnosis and treatment. Hospitals produce a large number of medical images every day, which are not only the major objective dependence for clinical diagnosis, illness tracking, prognosis research, and differential diagnosis for doctors, but also have significant value in diagnosis for the image that has not been confirmed in clinical. If we can find the images which have the same pathological performance and are globally similar with query medical image from the confirmed image database, it will greatly enhance the reliability of our clinical diagnosis. Therefore, image retrieval technology has been applied to the medical field.Because of the specificity of medical images, it is still an unresolved problem that how to organically integrate image retrieval technology with medical images so as to provides physicians with a favorable and convenient tool to retrieve images, and aid for diagnosis.Our research was carried out under the key projects supported by national natural science foundation of China (No:30730036). This thesis tentatively designs a brain image retrieval system, which focuses on retrieving images that have the same pathological performance, rather than globally similar content with the query medical image from image database, in order to help doctors to make disease diagnosis. The image database includes several modalities and three kinds of diseases. This thesis also explores two important parts of the brain image retrieval system—meticulous classification of brain images and automatic localization of disease area in brain images. Meticulous classification of brain images is applied to classify the images in database and query medical image. Retrieval performs only between query image and images from certain classes in database, so as to reduce the search space and enhance the retrieval efficiency. Automatic localization.of disease area in brain images should be generally effectual to most brain images, thus saving the doctors'labor on manual segmentation. The followings are the work of this thesis:1. Proposing meticulous classification of brain MR images based on support vector machine (ChapterⅢ)Brain images in typical MR-T1 scanning modalities are defined into 14 categories according to the difference on anatomical structure and content of images in a sequence of axial brain images. Firstly, selecting the training samples and testing samples, sample images are filtered by a smoothing filter and the backgrounds are removed. Use texture and shape feature jointly to express images, and then apply statistical association rule miner (StARMiner) algorithm to compute the weight coefficient of each feature. A meticulous classifier of brain images based on support vector machine (SVM) is trained, the parameters of which are optimized via repeating experiments. Rough classification rate is used in retrieval processing. Meticulous classification can be applied in special body part retrieval system in order to retrieve more accurate images and reduce the computational load and retrieval space.2. Proposing automatic localization and segmentation of lesion area in brain images (ChapterⅣ)It applies Hotelling transform to find the symmetry axis and rotate the image to the vertical position. Because normal brain images are almost of bilateral symmetry, we quantify the asymmetry caused by diseases and automatically locate the lesions in pathologic brain images. Segmentation only acts on the location and it can enable us to obtain the lesion boundary accurately. This method is fast and does not require the template image. It overcomes the shortcomings of poor real-time and the need of normal template image in registration-based segmentation method. Experiments have been performed on images of different scanning modalities and different diseases and it proves very effective.3. Designing a brain image retrieval system (ChapterⅤ)Design a multi-modality image retrieval system of brain diseases. Firstly, establish the DICOM image database and image feature database. The image database includes several modalities—CT and MR, and three kinds of diseases—cerebral hemorrhage, meningioma, and pituitary tumor. We design respectively automatic and semi-automatic segmentation of brain organization and localization of lesion area, and automatic extraction of the global and local features. We build a user interface of retrieval system including basic retrieval functions. Before features similarity comparison, computer selects the eligible sub-database from database, the images from which have the same position, modality, orientation as query medical image. Features similarity comparison is performed between query medical image and the images from the eligible sub-database, and then computer returns the retrieval images., Precision and recall curve are used to evaluate the retrieval system.