High speed networking is a crucial ingredient in medical information systems. ATM (Asynchronous Transfer Mode) and FDDI (Fiber Distributed Data Interface) networking have overcome their high costs and it is now possible to deploy these technologies widely. However, the impact of these networks on the performance of medical information systems is not well documented. Structured experiments are required to evaluate networking in complex systems of hardware and software. A test system designed to simulate an image delivery and display system using Ethernet, FDDI and ATM networking was established. Bottlenecks in this system related to networking protocols and hardware, as well as operating system and disk operations were identified and examined. Special attention was given to the DICOM (Digital Imaging and Communications in Medicine) protocol layered on TCP/IP (Transmission Control Protocol/Internet Protocol). Although the test ATM network was rated at 155 Mbps at the physical layer, observed maximum throughputs using TCP with optimized parameters were 79 Mbps for memory to memory transfers and 50 Mbps for disk to memory transfers. The default parameters yielded much lower throughputs with rates of 28 Mbps for memory to memory transfers. When the DICOM protocol was layered on top of TCP, a performance degradation of 10% to 78% of the optimal TCP rate was observed, depending on the type of image study being transferred (CR, CT, or MR). These initial data indicate the performance of a medical information system can be limited by a series of factors. Image data types have performance characteristics based on their image and study size. Appropriate selection and tuning of higher level protocols also makes a substantial contribution to system performance. Once network bandwidth exceeds ethernet speeds, disk operations are rate limiting factors in image retrieval.