Reference image databases begin to gain support

The development of medical image databases has been largely a private activity to date, lacking the rigor and standards that would allow these repositories to serve as reference models for research, including drug development. This is changing, however, as more organizations, including the U.S. National Institutes of Health and the European Federation for Medical Informatics, join the effort. But significant challenges remain, speakers said at a special session Saturday.

The incredible variety of imaging findings has proved to be an obstacle as researchers struggle to annotate reference images appropriately. A group at the NIH used a consensus process to annotate images with lung nodules but found it slow going, said Laurence P. Clarke, Ph.D., of the institute's cancer imaging program.

Resources have also proved to be a problem, according to Alexander Horsch, Ph.D., of the University of Munich and a member of the European Federation for Medical Informatics (EFMI). The federation established a workgroup in 2002 to pursue a reference image database. Since then, awareness of the need for such a database has grown, but the commitment to contribute has been limited.

"Everyone wants to get, nobody wants to give," Horsch said.

Two factors are driving the interest in reference image databases: the need for standards against which image processing strategies can be evaluated, and the use of imaging as an end point in drug development.

The benefits of a reference image database could be considerable, Clarke said. They include benchmarking for common cases, accelerated development of computer-assisted detection, objective assessment of therapies, greater confidence in clinical decisions, streamlined drug development, and better informed coverage decisions.

The National Cancer Institute launched its Lung Image Database Consortium project in 2001. This year, it reached a cooperative agreement with the American College of Radiology Imaging Network to share some of the clinical trial data. It has begun to speed up data collection and expects to complete the project with 400 cases in September 2006, Clarke said.

The database is intended to permit benchmarking of CAD for lung nodule detection and diagnosis in a screening and early diagnosis context. A recently established public-private partnership involving the NCI, FDA, industry, and the Foundation for the National Institutes of Health will build on the results.

Research by the EFMI working group identified other databases under development. The 10 largest include eight commercial databases, one with 1.2 million images. They generally could not be validated, however, and would not serve as reference databases.

A group of 10 smaller medical image databases comes closer to what the EFMI is seeking, Horsch said. One is the NCI's lung image database. Two are devoted to breast cancer: the Digital Database for Screening Mammography (U.S.) and one run by the Mammographic Imaging Analysis Society (U.K.). Three are committed to brain imaging: the Simulated Brain Database, the Biomedical Informatics Research Network, and the Laboratory of Neuro Imaging (all in the U.S.). The Pap Smear Tutorial (Denmark) addresses cervical cancer, and the Medical Image Reference Center (Japan) focuses on cancer and cardiovascular diseases.