IRMA A content-based approach to Image Retrieval in Medical Applications

INTRODUCTION Digital imaging has become the most prominent modality in medicine. Beside computed to-mography (CT) or nuclear medicine, which have been always digital, plain radiography, en-doscopy, and microscopy are nowadays captured directly digital, too. Picture archiving and communication systems (PACS) have been established in the hospitals all over the world to manage these information resources. PACS are used to store and handle the images, which are transferred based on the digital imaging and communications in medicine (DICOM) protocol. DICOM supports interconnections of PACS modules from different vendors, e.g. for post-processing of the images and their annotation with alphanumerical attributes such as patient and study information, image descriptions, and diagnostic reports. This textual information, which is stored within the DICOM header, is currently the only means to access and retrieve medical images from the PACS archive. Since an image tells more than a thousand words, recall and precision of this type of medical image information retrieval is limited in general [1,2]. This paper presents intermediate results of the IRMA project (http://irma-project.org) for content based image retrieval in medical applications. The IRMA project aims at describing medical images by means of their visual properties in an adaptive multi-resolution approach [3]. METHODS In a first step of processing, the images are categorized according to a terminology consisting of four orthogonal axis covering the anatomy (A) and bio-system (B) shown in the image as well as the creation (C) and direction (D) of imaging [4]. This is done based on global image features, i.e. a feature vector is assigned to the entire image. For each category, a global prototype is defined and used for both geometry and contrast registration of the images. Relying on a reference database of more than 15,000 radiographs that have been ABCD-classified by trained radiologists, an unseen image is categorized automatically. Currently, this categoriza-tion can be performed with an error rate of about 15%, 9%, or less than 5%, if the best match or a set of the five or ten best matches is considered, respectively [5]. The next part in the chain of processing extracts local image features. In contrast to the previous stage, a feature vector is now assigned to every pixel within an image. Based on an initial watershed segmentation, a region merging scheme is applied until the entire image is represented by a single region [6]. This final stage is considered as the root node of …