The Design And Implementation Of A Medical Image Retrieval Algorithm Based On Feature Fusion

With the development of information technology and medicine, CT, MRI and other digital medical equipment have been increasingly widely used in clinical work and auxiliary diagnosis. The modern equipment makes medical institutions produce large amounts of medical imaging data every day. As an indispensable part of the clinical diagnosis, medical imaging not only reflects the body’s anatomy and morphology, but also intuitively explains the function of human tissues and organs, which is an important objective basis for doctors to conduct clinical diagnosis, condition monitoring, surgical reference and pathological study.In the process of clinical diagnosis, doctors usually need to analysis and compare the images of different patients or the same patient at different times, then combined with the diagnostic report to make the final diagnostic conclusions. Doctors generally make judgments by comparing some local features of the same parts of the image, which need to be carried out according to the local feature retrieval, but a single local feature retrieve may appear erroneous detection of not the same portion. In some cases, local tissue features of different parts can be similar to each other, such as, certain lesions local features of brain images may be similar to that of the legs. So filtering input images of different organizations before retrieve is needed. To solve this problem, a medical image retrieval algorithm based on feature fusion is designed in this thesis.In this thesis, a medical retrieval algorithm which fuses global feature and local feature is designed. Firstly, use the global feature to make initial classification of medical image. Secondly, use the local feature to retrieve. SIFT feature points are extracted as local feature from image, and retrieve the local feature of images by constructing BOVW model. Specifically, a k-means clustering method is required to cluster the local features to get BOVW model dictionary. After that, use the frequency of feature words in a dictionary statistical image to describe image. Then index the image database based on the description vector of the image. Local features similar to other parts of the image are excluded by the pre-classification of the image. Specifically, a Tamura texture and a Context shape features Shape is fused to train the SVM classifier, then use the classifier to pre-classify image. Experimental results show that the proposed medical image retrieval algorithm based on feature fusion performs better than the traditional image retrieval method.