One Swedish and two Danish hospitals designed and validated a Web based tool for the automated interpretation of diagnostic heart images, based on image processing techniques, artificial neural networks, and vast medical databases. Dr. Lars Edenbrandt from Lund University introduced this AIDI-Heart project to the ITIS-ITAB'99 audience. The physician can use the computer-based service as a decision support assistant in cases where a more experienced medical expert is not available or as a professional means to obtain a second opinion for his own diagnosis. The user-friendly as well as accurate alternative for personal advice among medical colleagues, has the great advantage of being accessible 24 hours per day and 365 days per year.
The Internet-enabled user can start the application via a simple WWW-browser after he inserted a user name and the password. A common gateway interface (CGI) script is launched to send the local raw image data to the remote server for manipulation. The diagnostic advice is delivered within seconds after the image transmission. The system also allows the physician to return feedback with regard to the evaluation result. The collection and storage of this feedback on the server allows the AIDI-Heart team to optimize the decision support tool and still reduce the number of misinterpretations. Images are sent over the Internet in an encrypted state as to prevent patient data from being read by a third party.
In fact the AIDI-Heart service is used to detect coronary artery disease (CAD). The blood perfusion behaviour can be examined with the heart in rest and in stress on scintigraphic images that are acquired with a gamma camera, after injecting the patient with a radio-active tracer. The physician has the task to recognize the pattern in two so-called bull's-eye images where the dark parts represent areas with reduced blood flow. This type of diagnostics requires a large experience from the part of the physician. The availability of extensive medical databases of processed images turns artificial neural networks into high performing tools. Previous studies have indicated that neural networks can interpret heart images as good as an experienced physician.
The network has been tested both in Sweden and in Denmark to prove that the system is equally accurate in other hospitals than the one in which the application has been developed. The AIDI-Heart service also minimizes the variability of interpretations among different experts. The team has trained two sets of neural networks to detect CAD in the LAD territory of the heart, and in the RCA/LCX area. The output values range from 0 to 1 while three thresholds transform them into four different statements for the diagnostic advice. The architecture of the server, physically located in Lund, consists of HTML pages for login and demonstration functions, CGI scripts that manage requests, initialize additional programmes and store feedback, and external programmes for image manipulation.
Future plans of the AIDI-Heart team will involve the optimization of the user interface and the sensitive data security. Dr. Edenbrandt aims the service to be evaluated within a representative European user group as to demonstrate that intelligent information systems for image processing as well as decision support are very suitable for widespread use in hospitals. In this regard, the service need not be limited to CAD diagnosis but might be applied in other medical domains as well. The availability of qualitative medical databases of course remains of capital importance for the building of a reliable intelligent decision support tool. You can visit the AIDI-Heart home page for more news on the project.