In this study a dynamic, grid-based workflow is created to allow ad hoc composition of knee CAD algorithms in order to improve the performance of the overall CAD system.

There is significant, and largely untapped, potential for improving the overall performance of computer-aided detection by combining the results from different CAD algorithms. This can be clearly seen with CAD algorithms dealing with knee injuries. Knee injuries can occur with concurrent meniscus and articular cartilage injuries due to structural coupling. Existing CAD algorithms focus on the two processes separately. The performance of knee pathology CAD can be improved by leveraging knowledge of this coupling and correlating the output of both CAD systems. The purpose of this project was to facilitate such integration by combining multiple CAD algorithms for musculoskeletal imaging through a novel use of grid computing and grid enabled workflows.

The caBIG™ In Vivo Imaging Middleware includes DICOM data services for distributed image sharing and tools for creating grid-enabled, remotely executable CAD services. caGrid provides the middleware to create and orchestrate the workflow of CAD and DICOM grid services that integrates the results of the analysis performed by the multiple algorithms.[1,2] In our study, two CAD systems, one for detecting meniscal tears, and another for detecting articular cartilage injuries, were deployed remotely as CAD-grid services. Twenty-five MRI examinations of the knee were obtained utilizing our standard imaging protocol of the knee. The images were de-identified and then stored in a DICOM grid service. A radiologist then executed the workflow, resulting in images being retrieved from the DICOM service to the 2 CAD services for processing in parallel. The results of the CAD systems were sent to an aggregating result service, which combined them into a unified report for review and further analysis by the radiologist. In summary, the middleware coordinates data movement among the different services, manages the invocation of the algorithms and manages the subsequent collection of results.

This implementation demonstrates the advantages of leveraging grid technologies to integrate automatic detection of complex meniscal tears and articular cartilage injuries into a unified CAD system. Performance of the system was robust, and the automation of the process made it suitable for routine-preprocessing of MR examinations of the knee. These could be combined into a seamless workflow with resulting decision support information for the radiologist within a clinically acceptable timeframe.

Grid computing simplifies the rapid prototyping, deployment and integration of data processing components for image enhancement, analysis, and CAD. Using a standardized grid protocol, researchers can provide these components as grid services and explore and use other researchers’ grid services in image processing/CAD workflows. The ad hoc nature of the grid-based workflow allows rapid prototyping of improved image processing/CAD systems. caGrid provides an excellent platform for accessing and integrating the CAD algorithms from different institutions both as a research tool and for routine clinical use. Our study successfully demonstrated the ability to rapidly combine information for decision support and presentation to a radiologist or clinician from multiple algorithms located on different servers. The implications of this ability to send a single dataset to multiple servers in different locations are intriguing. The proposed system would enable the transmission of datasets to be processed in parallel by any number of available servers potentially utilizing multiple different approaches and then report the results in a unified manner, possibly combining human evaluation of the images as well. This could also enable small vendors with innovative algorithms to make their software available on the grid to the medical imaging community. The major drawback of the approach is that it requires data to be sent out of an imaging facility, which can be time-intensive, depending on the size of the dataset and may require de-identification of the data from a HIPAA perspective clinically or IRB perspective for research purposes.