Presented at the 1999 APIII Conference Return to 1999 Abstract Index
A CONTENT-BASED IMAGE RETRIEVAL SYSTEM THAT INTEGRATES A THREE-TIER STRUCTURE
University
of Pittsburgh School of Medicine
Department of Pathology
Pittsburgh, Pennsylvania
Lei Zheng, MS
Arthur W. Wetzel1, Lei Zheng2, Michael J. Becich, MD, PhD2
1Pittsburgh Supercomputing Center
2Department of Pathology, University of Pittsburgh School
of Medicine
Pathology practice heavily relies on image content. The
traditional glass slide is inconvenient for either transportation,
or archiving. Digital image database is a trend that addresses
these issues and also provides us with other potential applications.
Content-based image retrieval is a technology that allows
us to index a digital image database automatically based
on the feature of the image itself, employing techniques
from image processing, pattern recognition, artificial intelligence,
and database, so that the image database can be searched
with unknown images as queries.
We have been developing a content-based image retrieval accessing strategy since the very early stage of our digital pathology image archive innovation. A sample archive of about 600 images has been accumulated and used for algorithm developing and testing. The image search engine residing on a supercomputer, does intentive computation such as image processing, feature extraction, comparison, and database accessing. To access the relatively secured image archive and supercomputer, and provide the user with an easy access method and friendly interface, a simple, intuitive, light-weight graphical user interface is developed as a separate client component which can be run on the user's machine using Java technique. It provides networking capability, receives user input, and delivers query results. A middle-ware is also under development to bridge the supercomputer and the Java client to add extra features and flexibility to future scalability.
The primary system has been tested successfully at multiple locations with different network setup, and for various audience groups. We are currently working on improving all three parts of the system to reach a higher overall retrieval accuracy, and better usability.
