film digitizer, and saved as a grayscale TIFF file with resolution 2K x 2.5K x 12 bits. The average TIFF file size is approximately 10 megabytes. TheTIFF image is ...
A Web-Enabled Research Database with Image Recognition Ronda Dulbandzhyan, Raymond G. Duncan, M.D., and David L. Rimoin, M.D., Ph.D. Departments of Enterprise Information Services and Pediatrics and the International Skeletal Dysplasia Registry Cedars-Sinai Health System, Los Angeles, California PhotoshopT and re-saved as a highly-compressed JPEG file which has a typical file size of less than 250K. The JPEG files are then loaded into the Oracle database as a "BLOB" data type. Each patient may have from zero to many of these JPEG files linked to his/her record as needed. Each JPEG is associated with text data elements including diagnosis, patient age at time of X-ray, Xray procedure, X-ray view and original JPEG file name. These elements support a Knowledge-Based Retrieval Approach [1], allowing investigators to search for and view images based on their labeling. But investigators also wanted to be able to perform searches based on image characteristics. Oracle's Visual Information Retrieval. (VIR) product was evaluated as a solution to this requirement. VIR analyzes the original TIFF image to extract key visual attributes, including Global Color, Local Color, Texture, and Structure. These attributes are encoded as a vector that is stored in the database alongside the compressed JPEG image. To retrieve similar images with VIR, users can essentially query by example, starting with a representative X-ray and assigning a weight to each visual attribute. These vector-based VIR searches can be combined with criteria based on the diagnosis, procedure, view, and age to rapidly isolate a few Xrays out of many thousands [2]. The attribute threshholds that determine whether two images are considered to be similar are also under user control. Discussion Precipitous decreases in the cost of processing power and disk storage enabled the rapid migration of an existing research database to a modem 3-tier architecture with a web user interface and the incorporation of large numbers of scanned images. Current enhancement efforts are focused on the integration of Oracle VIR technology to support retrieval of patient records and X-ray images based on visual characteristics. References [1] Lowe, H. et al; Towards Knowledge-Based Retrieval of Medical Images; Proc AMIA Annual Symposium 1998; 882-885 [2] Oracle8i Visual Information Retrieval, Data Sheet, Oracle Corporation, February 1999; 2-5 Acknowledgements Supported in part by a NIH Program Project Grant
Abstract Dr. Rodney Roentgen is studying a series offilms for a new patient with achondroplasia. The pelvis is typical of the disorder but there are some unusual features in the hands. Dr. Roentgen would like to determine if there are X-ray images on file with these features for any other patient that has been seen by the International Skeletal Dysplasia Registry. Using the traditional approach, Dr. Roentgen would have to manually inspect thousands of other X-ray films to try and find a match. With the use of stored database images, Knowledge-Based Retrieval data elements and the Oracle Visual Information Retrieval (VIR) cartridge, this searching process can be greatly streamlined, Background The International Skeletal Dysplasia Registry (ISDR) was first established in 1968. It serves as a world-wide referral and diagnostic service for patients with skeletal dysplasias. The registry consists of over 8,000 cases and an archive of 50,000+ X-ray images. On-going research is funded by NIH grants. The current active research program includes biochemical and DNA studies. In 1988, a database was developed in Nantucket Clipper. to store patient registry numbers, patient demographics, associated physician and health care facility information. In 1998, the Clipper database was reversed engineered and converted to an Oracle 8.0 database. The existing functionality and data was migrated to the new system, image storage was added, and the user interface was web-enabled using PL/SQL program units, HTML, and the Oracle Application Server version 4.0.7. The decision to web-enable the new database was based on minimizing costs for training and support, the need for platform-neutrality and secure remote access, and the desire to integrate images with the demographic, clinical, and research data. Methods Textual patient data is entered and updated using a series of web pages and forms. Frames are used to maintain separate menu and data entry areas. Keyboard requirements are kept to a minimum by means of pick lists and menus. In order to add an image to the database, the Xray is scanned using a Lumisys Lumiscan m 75 laser film digitizer, and saved as a grayscale TIFF file with resolution 2K x 2.5K x 12 bits. The average TIFF file size is approximately 10 megabytes. The TIFF image is then balanced and cropped with Adobe
1091-8280/99/$5.00 C 1999 AMIA, Inc.
(HD 22657).
1052
