The purpose of this study is to develop and evaluate the utility of an alternate approach to transferring patient data between facilities by deploying CAD algorithms to these facilities.

caBIG™ is a national effort to create an infrastructure which enables sharing of data and applications, and promotes collaboration in the cancer research community. However, caGrid, the tooling that enables this collaboration, is domain agnostic and can be used in non-cancer research as well.[1,2] In addition to technological challenges, there are various privacy and intellectual property issues that arise when sharing data across institutions. Protecting PHI data has always been a major issue when technologies that foster collaboration and data sharing are involved. It becomes a challenge, not only for IRBs, but also for institutional IT staff when attempting to form such collaborations. Institutional IT staff and IRB’s have struggled with potential security and privacy compromises when images and related data are transferred outside an institutional firewall, increasing the risk of the data being compromised. Our aim is to demonstrate a new paradigm which addresses these privacy concerns by transferring an algorithm to the site that contains the data that is to be processed, rather than the traditional approach of sending data to a server running that algorithm, thereby ensuring that PHI data never leaves the hosting institution. To our knowledge there are no IRB issues that relate to the downloading of algorithms for local use on a dataset as long as they are checked for viruses and meet with the approval of local IT and radiology policies.

The NCI caBIG™ In Vivo Imaging Middleware has developed tools that leverage grid technologies to accomplish this proposed on-demand dynamic deployment of algorithms. The system has 3 components: (1) A grid data service that stores 25 MRI examinations of the Knee; (2) A remote analytical grid service that stores a MATLAB Knee CAD algorithm, developed by the University of Maryland School of Medicine (UMSOM); and (3) The dynamic service deployment middleware that orchestrates the on-demand deployment and subsequent execution of the algorithm. For this study a radiologist requests the Middleware for a CAD deployment. Upon receiving a request from a client side workstation the middleware responds by transferring the CAD algorithm from UMSOM to a remote algorithm repository grid service that is collocated with the image data service. The middleware then executes the algorithm on the set of images identified in the original request. Upon successful execution, the results are sent to the radiologist. The middleware also secures the intellectual property of the algorithm, by deleting the algorithm after it has successfully completed its execution.

A pilot system consisting of a CAD grid service and an image data service running at a different site was deployed. Radiologists were able to issue requests for the algorithm to be deployed and executed at the site of the image service. After successful execution, the results were sent back to the radiologist and the algorithm was undeployed. The time required to download, utilize, and delete the algorithm was relatively short and did not adversely affect image interpretation workflow.

Privacy has always been a concern when grid services are used in research, especially when data is transferred beyond the firewall of a research or clinical facility. Our solution offers the potential to alleviate this problem. This novel strategy of moving an algorithm to the site where the data to be processed is located, has two advantages. First, data privacy and integrity concerns that arise in multi-institutional collaborations can be significantly reduced, since the data never leaves the hosting institution. Second, this approach reduces the overall execution time, as the cost of transferring an algorithm is typically smaller than that of transferring data. Additionally, this approach offers the potential for a business model in which a workstation could have access to multiple proprietary or commercial algorithms, as well as freely available algorithms, as long as they share a common framework such as consistency with the specifications promoted by DICOM working group 23.

Multi-institutional collaborations require careful planning to protect patient privacy and data integrity. This new paradigm of using grid technologies and dynamic service deployment will address and may, in the future, significantly alleviate concerns of data privacy that arise in these collaborative environments. This study illustrates a novel approach to improving data privacy in a shared research environment. Additionally, since the size of image datasets is significantly greater than algorithms, there is also a secondary advantage of reducing the overall execution time, since the images are never transmitted. This approach also offers some intriguing possibilities for a business model within a single vendor or even using the software provided by multiple vendors on a workstation or using an Internet “portal” approach.