| With the appearance of medical imaging equipment in the medical field, medical imaging has become a basis tools for modern medical, there will produce a large amount of medical images which contain physiology, pathology and anatomical information in the hospital every day. It plays an irreplaceable role in clinical diagnosis and treatment, medical teaching and research, which produces a huge demand for effective management and retrieval to the medical image data. Traditional image information retrieval rely on text annotation and keyword which brings some serious problems:on the one hand, now the image information data is very large, we can not achieve the standard by manual review and comment on images; on the other hand, human observations and comments are very subjective, different people may have different views and comments to the same picture, and even the same person would give different evaluation at different times, which is very unreliable. Thus, the content-based image retrieval techniques became a research focus in the last decade, but also are an effective way to solve the massive image data management and query.Content-based Medical Image Retrieval (CBMIR) extracts gray, shape, texture, other low-level visual features and high-level semantic features from image itself, constitutes a feature vector to describe the image content, indexing and matching criteria with feature vectors to retrieve the image. In recent years, CBMIR technology has become a very active field of research in biomedical engineering.Currently, CBMIR technology is mostly from content-based image retrieval (CBIR), but in CBIR many mature technologies can not be directly transplanted to CBMIR, which is mainly decided by the characteristics of medical image itself:1) Most medical images are grayscale, different images may have same grayscale, which makes the conventional image-based global features (such as the gray histogram) does not apply to medical image; 2) Natural images are often contain the background color which can be distinguished, however, the background of medical images (such as X-ray images) are black, the distinction of image category is only relying on the classification of objects in the image; 3) There is a variety of medical imaging modes, the image processing algorithms and feature extraction will be different; 4) Image processing technology is still confronted with many difficulties in medical image processing, such as medical image segmentation; 5) Medical imaging process will introduce all kinds of noise, artifact and geometric distortion interference; 6) The grayscale resolution and spatial resolution are high in medical image, which contains a large amount of information. Therefore, we must seek for feature extraction and content representation which suit for the characteristics of medical image, to make CBMIR success.The basic steps of Content-based Medical Image Retrieval are:do image preprocessing and analysis to all images in the image database, extract image feature, establish signature, associate signature with image library with a specific identity; in the retrieval process, we first extract the feature vector of query image, and then compare with the feature vectors in the signature, according to the matching results in the image database to retrieve the required image to return to the user; However, the two images with similar underlying characteristics can vary greatly in the semantic content, there is the so-called "semantic gap", which is the main reason to restrict the properties of the CBMIR.The causes of semantic gap are:1) There is a certain gap between computer vision expressed and human perception of color, general similarity of the distance is just the distance between the image feature space, rather than the real similarity distance between image semantics; 2) The gap between high-level semantic and low-level features, people are always use some high-level concept in their daily lives, but the features extract from images are mainly low-level features, in addition to particular areas such as face recognition and fingerprint recognition, in most cases, it is difficult to directly establish relation between low-level features and high-level features; 3) The subjectivity of human perception. For the same visual content, different people or the same person might have a different understanding under different circumstances, this is the subjectivity of human perception. This subjectivity may exist on different levels, for example, someone might be interested in the color of the image; while the other may be more emphasis on the texture characteristics; or both are equally focus on texture, but the knowing of texture may be completely different. In fact, the expression of any form of texture features can not cover the complete image information, a different feature representation reflects the visual characteristics from a different angle; 4) There are some image retrieval systems using a multi-feature retrieval, due to different features have different similarity measures, it is difficult to find a more suitable distance to meet the human perception.How to solve the semantic gap is one of the most active issue in CBMIR research. There are some technologies to reduce the "semantic gap":region-based image retrieval, image semantic classification based on machine learning, relevance feedback(RF), adaptive similarity matching function, CBMIR combined with the text etc, RF is the most widely used techniques and its results are obvious, RF allows users to evaluate and mark the search results, and give feedback to the system, then the system uses the feedback for learning to return query results which are more in line with user requirements.Relevance feedback was originally a retrieval system used in the text, which uses the previous search results to automatically adjust the current query. In the content-based image retrieval, the search is always done through a series of interactive. First, the user does an initial query, the system returns the closest results of the query. Then, the user evaluates the feedback results, marks satisfactory results as positive-examples, and marks dissatisfied results as counter-examples. The system does self-adjustment and a new round of retrieval according to the marked results. Simply, that is, we take the results of the previous query to the system's feedback to improve the next output. The process can be repeated according to user requirements. The objective of relevance feedback is learning during the process of actually interact between the users and the query system, discovering and capturing the actual user query intention, then correcting system's query strategy, in order to obtain the results consistent with the actual needs of the user as possible. Relevance feedback can modify query strategy in real time, so it increases the adaptive function for image retrieval system.This paper first analyzes the research status and existing problems of content-based image retrieval, detailed the various key technologies in content-based image retrieval, including the development stage of image retrieval technology, the evaluation and analysis of retrieval performance, the extraction and expression of image visual features, etc, and analysis relevance feedback in a variety of algorithms. Details the applications of nonnegative matrix factorization (NMF) relevance feedback to medical image retrieval. Traditional multiplicative update NMF approach has so many iterations and slow convergence time, so we proposed projected gradient approach to the NMF, in the assurance of precision ratio and recall ratio, it can greatly improve the speed of the retrieval. This article includes the following tasks:1) We system introduce the basic knowledge and key technologies for content-based image retrieval, for the characteristics of medical images, propose feature extraction suit for the gray, texture, shape, space, detailed analyze some common feature extraction methods, integrate various features to achieve greater efficiency of retrieval.2) In order to reduce the gap between low-level image features and image content in the semantic, which is "semantic gap", we introduce the relevance feedback technology into the image retrieval system, based on the analysis of all relevant feedback, we propose NMF (non-negative Matrix decomposition) method, traditional multiplicative update NMF approach has so many iterations and slow convergence time, so we proposed projected gradient approach to the NMF, experiments show that in the assurance of precision ratio and recall ratio, it can greatly improve the speed of the retrieval. |
PDF Full Text Request