The Multi-Media Database Project

Prof. Ehud Gudes
Department of Mathematics and Computer Science
Ben-Gurion University of the Negev
Beer-Sheva, 84-105
ehud@cs.bgu.ac.il

The main goal of the research group headed by Ehud Gudes is the development of the Multi-media database and the Software foundation required for the other researchers in the project to develop and evaluate their respective image/voice processing techniques (Compression, Restoration, Segmentation, etc.), and the development of the Database system needed to store the Medical application data and the Images/Audio information. The second goal is to investigate important database issues in a dynamic and rich-structured environment such as Medical multi-media information system.

Multi-media database and Retrieval system

The database stores various types of multi-media information including textual, images and voice. Currently it is implemented in Java on top of the Informix relational DBMS. In the future we plan to port it to an Object-oriented or an Object-Relational database.

The retrieval system which uses the Multi-media database stores and manipulates Patient data in the Oncology department of the Soroka hospital. Currently, the system supports a database with 10 patients files including the following types of information:
- demographic items
- medical history
- present disease
- Body diagrams (see below)
- CT, XRay images
- Blood and biochemistry tests
- Speech fragments spoken by the treating physician

It is now possible to browse through a single patient file, or to query multiple patients files. The output may be textual information, images or played voice.

The system provides various processing tools on the above Multi-media database. The most important tools are:

• A compression/decompression tool based on the algorithms devised by Prof. Awerbuch. This is a lossy compression scheme which enables very deep compression, (ratio of up to 100 in size of resulted files). The algorithm operates on rectangles with sides which are power of 2 pixels. The interface enables the following:
  -It allows a user to define different compression rates for different parts of an image (by drawing the desired rectangles and specifying their compression rate). A user can easily divide an image into rectangles and define a different compression rate for each such rectangle, thus allowing non-important areas to use high compression, and very important areas to use low compression (and low loss).
  - The interface allows to display a compressed image, by applying a method that joins all compressed files (i.e. the corresponding Rectangles).
• The drawing and manipulation of Body diagrams which contain very helpful information for the physician concerning the location, the size, and the type of the tumor. This includes:
  - A collection of images with different body parts.
  - A user can choose any of these images and mark on it the region to be treated by radiation. The user may also write some necessary textual information on the image.
• The processing of CT and X-ray images by an expert and the search of objects and object features afterwards. This module supervised by Prof. Dinstein currently supports:
  - the identification of a region suspected as "tumor", and the "coloring" of its internal surface or edges.
  - the search for images with similar "tumors"
• The search of content of continuous speech fragments for a given keyword. (KWS - Keyword spotting). In particular since diagnostics of patients are often represented as Speech files, it allows the search of these files using pre-defined keyword. This project is supervised by Prof. A. Cohen. It allows the following:
  - The training and derivation of an HMM (hidden Markov) model for a given keyword from a set of keyword speech files
  - It allows the computation of a threshold value which distinguish between the keyword and similarly sound but different words ("garbage")
  - It uses the HMM model and the Threshold function in searching for a given keyword in a set of continuous speech fragments which are part of the patients database.
  Thus with additional training the capabilities of this search mechanism are increased.

From the Architecture point of view, the system consists of two parts: interface and server.

• The interface sends to the server messages to fetch data from the database, and presents and processes the received data. The Interface is written in Java.
• The server retrieves data from the database by sending queries, that it received from interface, to Informix. The fetched data is sent to the interface. The server is written in C with embedded SQL.

The above architecture allows us to put the user-interface on a remote machine and thus providing a Remote access.

Database Research topics
The following topics are related to our Multi-media database research:

1. The investigation of Dynamic indexing techniques. A new scheme for dynamic and adaptable indexing was developed. This scheme provides Indexes which change their behaviour based on the User's queries, thus they avoid the difficult problem of Index selection. Details of this research can be found in [1].
2. The investigation of Propagation semantics in Object-oriented databases. A new algorithm which is based on the original algorithm by Rambauch was developed and applied to a complex object-oriented database. Details of this work can be found in [2].

References:

1. A. Raisman, A B-tree based adaptable and dynamic Indexing scheme, MSc thesis, Ben-Gurion university, 1998.
2. A. Sapir, Propagation semantics in object-oriented databases, MSc thesis, Ben-Gurion university, 1998.

• About this document ...