Naveen Garg MD, Presenter: Nothing to Disclose

David Joseph Vining MD, Abstract Co-Author: Royalties, Bracco Group

David Fuentes, Abstract Co-Author: Nothing to Disclose

Kevin W. McEnery MD, Abstract Co-Author: Medical Advisory Board, Research In Motion Limited Advisor,Koninklijke Philips Electronics NV

John D. Hazle PhD, Abstract Co-Author: Nothing to Disclose

Vikas Kundra MD, PhD, Abstract Co-Author: Nothing to Disclose

The use of automated image processing and analysis of multiple prior examinations is possible in a PACS environment using a cluster of minicomputers to accelerate data transfer . 

The increasing number of radiologic examinations and the number of images per exam threatens to outpace Moore’s law which anticipates a doubling of computing power every 18 months. While supercomputers are available at large academic and commercial enterprises, the infrastructure needed to reliably and efficiently transfer and process petabytes of data is still an unsolved issue. We present an image processing cluster, of headless minicomputers dedicated to image processing and decision support, that could even be used by smaller radiology practices.   

Fifteen Zotac-mag minicomputers were used in developing the cluster. Each computer has a 160 gigabyte hard drive, 2 gigabytes RAM, 1-gigabit Ethernet connection, and no monitor or peripherals devices. The cluster is backed by an 11-terabyte network storage device. As a radiologist opens an imaging study for review using a standard PACS system, the cluster loads individual prior examinations on each drone computer. A custom software application provides for feature extraction from each of the studies based on automated image analysis . Post-processed data are then made available for remote visualization over the network for presentation on the local PACS workstation.  

A proof of concept distributed image processing system was developed and tested using our testing PACS server environment. Use of multiple inexpensive minicomputers as servers with their own gigabit ethernet connections may allow for faster data transfer than having a single more powerful server with a single gigabit ethernet connection. Simultaneous maximum intensity projections (MIPS) of multiple prior CT examinations is prototyped. Algorithms to pre-fetch and cache anticipated imaging data are under further development to improve throughput. Future work in machine learning algorithms will be essential to assist the radiologist in assessing and gathering new information from the large amounts of imaging data collected from the multitude of exams.

Garg, N, Vining, D, Fuentes, D, McEnery, K, Hazle, J, Kundra, V, An Image Processing Cluster for Radiology Decision Support.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11010211.html