Knowledge-based Medical Image Retrieval

While the development of medical imaging technology promotes clinical medicine, digital medical images play an increasingly important role in Computer-Aided Diagnosis, teaching and biomedical researches. At the same time, the number of digital medical images is increasing rapidly, then, how to get the needed images from the large volume of image data becomes an urgent problem. However, neither traditional keyword-based image retrieval nor current content-based image retrieval can meet this need well. So, it is necessary to explore new methods to effectively index and retrieve medical images.This dissertation focuses on the approach which retrieves medical images based on their corresponding text, and explores knowledge-based medical image retrieval methods. The corresponding text of medical image includes DICOM header, medical reports, image annotations, etc., which contains information about the medical image. As the text information is in natural language form, they are more close to user's search habit and human's perception of image. Hence, it is more effective to retrieve medical image based on their corresponding text than based on their low-level features. Considering that there are many defects in traditional keyword-based text retrieval, for example, synonyms decreases the retrieval recall and polysemy lead to a low precision etc., this dissertation proposes a knowledge-based medical image retrieval method. Based on a medial ontology, medical concepts are extracted from the corresponding text and then used as metadata to annotate images. In this way, medical images are assigned semantic meaning and can be retrieved by concept matching on knowledge level other than by keyword matching on syntax level, so the synonyms and polysemy problems can be solved. Furthermore, knowledge-based medical image retrieval can use the parents and children relations of concepts to expand queries, and can achieve cross-language retrieval by using multi-language thesauruses.As the diversity forms of concept description in real-world medical reports and articles are rife, medical concepts can not be extracted accurately and completely with the traditional information extraction methods. To solve this problem, a new medical concept extraction method is proposed, and based on it, an exploratory research of indexing and retrieval method, as well as the design of retrieval framework are carried out. The main contributions of this dissertation are as follows:1. Considering the diversity forms of concept description in real-world medical reports and articles, a mixed concept extraction method is proposed. The method is consisted of two extraction models. One is Maximum Matching Model which is based on knowledge understanding, and another is Minimum Matching Model which is according to the composition analysis of medical terms. The merits and defects of the two models are discussed in detail, as well as the ways to improve the defects. It is expected that combining the two models can improve the performance of medical image retrieval.2. To achieve knowledge-based retrieval, classical Vector Space Model is replaced by Concept Vector Space Model. Concepts are weighted by TF-IDF weighting strategy, and similarity between queries and images are calculated by Cosine similarity measure.3. To verify the proposed method, several experiments are conducted on Image Based topics of ImageCLEFMed2009. The experimental results show that the mixed concept extraction method with queries expansion is an effective way to improve the precision of medical image retrieval.4. A knowledge-based medical image retrieval framework is proposed, which integrate the aforementioned methods. And a medical image retrieval prototype system is developed based on this framework. The system is developed with component technology, to achieve Low Coupling / High Cohesion and high scalability. In this system, various retrieval models can be easily integrated, and the system can also be easily integrated with other retrieval system. Several instances of system can also be combined to form a distributed retrieval system. All components of the system can be easily configured with xml files, and all configuration parameters can be adjusted in run-time.