Teleradiology offers significant improvement in efficiency and effectiveness over current practices in traditional film/screen-based diagnosis. In the context of digital mammography, the increasing number of women who need to be screened for breast cancer, including those in remote rural regions, make the advantages of teleradiology especially attractive for digital mammography. At the same time, the size and resolution of digital mammograms are among the most challenging to support in a cost effective teleradiology system. This paper describes a teleradiology architecture developed for use with digital mammography by GE Corporate Research and Development in collaboration with Massachusetts General Hospital under National Cancer Institute (NCI/NIH) grant number R01 CA60246-01. Experience with a testbed prototype is described. The telemammography architecture is intended to consist of a main mammography diagnostic site serving several remote screening sites. As patient exams become available, they are forwarded by an image server to the diagnostic site over a WAN communications link. A radiologist at the diagnostic site views a patient exam as it arrives, interprets it, and then relays a report back to the technician at the remote site. A secondary future scenario consists of mobile units which forward images to a remote site, which then forwards them to the main diagnostic site. The testbed architecture is based on the Digital Imaging and Communications in Medicine (DICOM) standard, created by the American College of Radiology (ACR) and National Electrical Manufacturers Association (NEMA). A specification of vendor-independent data formats and data transfer services for digital medical images, DICOM specifies a protocol suite starting at the application layer downward, including the TCP/IP layers. The current DICOM definition does not provide an information element that is specifically tailored to mammography, so we have used the DICOM secondary capture data format for the mammography images. In conclusion, experience with the testbed is described, as is performance analysis related to selection of network components needed to extend this architecture to clinical evaluation. Recommendations are made as to the critical areas for future work.