Peter C. P. Lau MD, Presenter: Nothing to Disclose

John Marshall Knudsen MD, Abstract Co-Author: Nothing to Disclose

Brian Jack Bartholmai MD, Abstract Co-Author: Nothing to Disclose

Timothy P. Maus MD, Abstract Co-Author: Nothing to Disclose

Bradley James Erickson MD, PhD, Abstract Co-Author: Nothing to Disclose

A significant and growing proportion of radiological examinations are 3D in nature—CT, MR, PET. There are computer algorithms that can reliably align images from one study with another—both across modalities and through time. The purpose of this study was to evaluate the impact of automated image alignment for routine imaging examinations on measurement reproducibility and interpretation time. We also looked at limitations and problems that might hinder routine clinical application.

After IRB approval, we recorded all examinations performed on CT and MR scanners over a 3 day period and the examinations that had relevant prior CT or MR examinations. The first 100 neuroradiological and 100 body cases with priors were enrolled. We retrieved both the ‘current’ exam and the most recent prior to a workstation. All studies were anonymized. Next, we manually identified corresponding series. We then used an ITK-based image alignment algorithm to align the images of the old exams onto the new exams. Both the unregistered pairs and registered pairs were transferred (as different patients) to a PACS test system. In cases where there was a significant finding, they were asked to measure the lesion. In all cases, they were asked to provide a confidence level of their findings. The time to reach this conclusion was also recorded. If the registration algorithm failed, that was also recorded.

72% of neuro cases and 55% of body cases had relevant priors. In 93% of neuro cases, the relevant prior exams could be aligned to within 1 pixel. The success rate for body cases was much lower, largely because chest and abdominal structures often had very different shapes do to physiologic body changes (breathing, etc). In those cases where alignment was satisfactory, there was a reduction of interpretation of approximately 10%, an increase in confidence level and an increase in measurement agreement.

Image registration can increase accuracy and efficiency, but some body parts are more amenable than others.

Lau, P, Knudsen, J, Bartholmai, B, Maus, T, Erickson, B, Impact of Automated Alignment on Interpretation of CT and MR Studies.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414392.html