Automated image processing bolsters radiology workflow

Although it's been around for more than a decade, multimodality image registration hasn't entered routine clinical use. But a team from the U.S. National Institutes of Health (NIH) believes an automated data processing system can clear workflow hurdles, sparking adoption of imaging registration and other advanced image analysis tools.

The NIH has developed what it calls an automated postprocessing system (APPS) that automatically performs tasks such as multimodality image registration and physiologic image processing and makes them available to radiologists as they read studies. The system has become invaluable to the institution's clinical workflow, said Jeffrey Solomon, PhD.

"Advanced image analysis is accessible to radiologists because this is fully automated," Solomon said.

Solomon shared the institute's experience with automated image processing during a session at last month's Society for Imaging Informatics in Medicine (SIIM) meeting in Orlando, FL. He teamed on the project with Dr. John Butman, PhD, a neuroradiologist with NIH.

Lagging adoption

Many advances in medical imaging analysis have not penetrated into clinical workflow, including multimodality image registration, dynamic image processing, and tumor segmentation and measurement, Solomon said.

For example, in evaluating tumor response to treatment, radiologists have the critical task of assessing the size of the tumor at one point of time and then later during follow-up to determine change. However, because of differences in patient position between scanning sessions, the 2D slices do not depict the same anatomy, he said.

The solution to this problem is image registration, which automatically aligns the scans. This facilitates visual inspection as well as quantitative measurements, Solomon said.

In standard workflow, radiologists use PACS to quickly access medical images and dictate reports. In an ongoing study sponsored by the U.S. National Cancer Institute (NCI), PACS is also being used to store and retrieve images for ongoing study of brain tumor therapies, he said.

To provide radiologists with access to these advanced tools, technologists were asked to transfer images from PACS to a third-party workstation, where they performed automated registration and perfusion processing. The results were then submitted back to PACS. As the number of patients increased to more than 2,000, however, a more efficient system was required, Solomon said.

Workflow integration

As a result, the group sought to develop a fully automated postprocessing system that would require no user interaction and would integrate image registration and physiologic image processing into the radiology department workflow. They also wanted to ensure compatibility with standards such as the Integrating the Healthcare Enterprise (IHE) initiative and the DICOM standard.

Making use of IHE's Postprocessing Workflow Profile, APPS queries a broker that ensures correct demographics on the imaging modalities, and creates an internal worklist just for the studies that will receive postprocessing, Solomon said.

Via a DICOM C-Move process, APPS then queries and retrieves images from the PACS when they're available, performs postprocessing, and sends the results back as an additional study to PACS.

"All of this communication is done without anybody touching any buttons," he said.

A radiologist will request images from the PACS viewer, get them from the archive, and dictate the report, he said.

APPS can currently perform multimodality image registration to an initial reference scan and handle MRI and PET images. It can also calculate PET standardized uptake values (SUV) and perform automatic windowing/leveling.

Other features in the modular system include dynamic susceptibility contrast perfusion processing via automated input function selection and deconvolution, coregistration of diffusion-tensor imaging, and automated DICOM printing of coregistered datasets on a single sheet of film for patient communication. A DICOM router can handle studies from outside institutions, Solomon said.

Sometimes failures occur when a study has incompletely transferred from a scanner, when wrong studies are sent manually to APPS, or in the rare event that a study has been incorrectly registered. As a result, the group developed a mechanism to reprocess data manually, he said.

The system uses FMRIB's Linear Image Registration Tool (FLIRT) (University of Oxford), the DCMTK DICOM toolkit (OFFIS) for DICOM communication, and MedX image analysis software (Medical Numerics). The team also developed homegrown C/C++ programs and shell scripts to tie the whole system together, he said.

Performance

So far, APPS has been used to analyze images from more than 20,000 exams of more than 2,000 patients enrolled in NIH clinical trials, Solomon said. Processing is run on a distributed Linux workstation with multiple processors.

Average processing times are as follows for different components of the software:

  • Registration engine: 3.41 minutes
  • Perfusion engine: 9.16 minutes
  • Printer engine: 15 seconds

APPS has become invaluable to the clinical workflow of the institution, making advanced image analysis accessible, Solomon said.

"That comes from all of the neuroradiologists who use it and rely on it, and also the referring physicians at NCI who are very happy with this," he said.

Logging systems and dashboards have been developed to monitor system performance, he noted.

Solomon and colleagues hope that adherence to international standards such as DICOM and IHE will allow the integration of APPS into other radiology departments.

The team is currently integrating susceptibility-weighted imaging (SWI) into APPS for use with a traumatic brain injury cohort. Other future possibilities include the integration of automated tumor segmentation methods and collaboration with other institutions, Solomon said.

Page 1 of 383
Next Page