Content-based Image Retrieval System For Medical Image Database

With the improvement of image-based clinical technique, large amounts of digital images are produced everyday. The management of medical image database and how to use those images in Clinical Diagnose Process (CDP) becomes a especially challenge in medical field. The most important thing to improve the efficiency of traditional PACS is to introduce the CBIR (Content based Image Retrieval) technique to the manage process of medical images. Traditional PACS use a handy (or semi-automatic) method to manage images in database. It not only runs with a very slow speed, but also reduces its clinical benefits in CDP. CBIR provides the ability to access images by their content, which meets the requirement of PACS. So, it is obvious that we should provide the query-by-content ability to PACS. But it is a pity that, although the need to migrate CBIR in medical images treatment has been proposed many times, it still in its junior stage。As the DICOM standard has been made for a long time, the need for remote diagnose became more and more popular, all of this challenge the access method of images in a rough manner. The DICOM standard also gives some convenience to CBIR, because the structure of DICOM image contains not only original image data, but many other information for use in the diagnose process. The motive of this article is exploring the usage of CBIR technique in management of medical images, and develops a CBIR system base on medical images that can improve the access ability of medical images. May those work will contribute to the implement process of CBIR in medical. First part of work of this article is on feature extraction method of images. From the partition-based histogram, we develop a new feature extracting method. Histogram is a classic method of expressing content of image, but because of its globe limits, it can not be applied to application directly. Partition-based histogram is the improvement of traditional histogram, which introduces some spatial information of image and made it possible to query images by its histogram features. In this part of work, we first introduce a new partition method based on circular split. This method has two main advantages: Firstly, partition according distance to image center matches the way our vision system works, and makes weight dispatch between partitions more reasonable; Secondly, refined partition based on circular split make it is possible to hold enough spatial information in image features. After image partition, we use a new feature extracting method on each block which composed of two components. First part of feature is main-color-histogram which only computes colors with large ratio; second part is a vector represents distances between self and all other blocks in this image. Main-color-histogram reduces dimension of image features which makes our algorithm runs faster and most important of all it removes the affect of histogram's "zero point". The present of distances between blocks makes the feature contains more spatial information which can improve performance. Our algorithm passed the common criterion's test, which identify our algorithm's efficiency. Texture is another important feature of image. In our system we use some texture parameters to cooperate our method with the image retrieval process. To improve our system's usability, we introduced the technique of relevance-feedback. As far as we know, the features adopted by most CBIR system contain color, texture, shape and relation-based shape which gained more attention these years. All of them are low-level features and differ from the way brain describe objects. So, the recall rate and precision rate of current CBIR system are all not very satisfacting. Relevance-feedback is the mostly used technique to refine the query result in CBIR system now. Since it was proposed by Recchio in 1971, it has been adopted by many information retrieval systems. It has been proved that the use of relevance-feedback method can improve system performance significantly. In our system, relevance-feedback is applied in two ways: re-weighting based and query point movement based. We have developed a platform for CBIR and implement our Medical Image Retrieval System from it. In the development of the CBIR platform, we employed Object-orient techniques to make it independent to the knowledge of field. Under the roles we have made, one can easily add additional features and matching rules. Our CBIR system for medical images is right developed from the platform by adding some such method on it. When CBIR is applied to specific field, the most important and most difficult problem is how to use the knowledge in the field. There are so many kinds of images that we can not expect to use a single pattern to distinguish them. Every kind of images has its own characteristics, this is especially true...