Effect of Automated Image Registration on Radiologist Interpretation

Abstract

In this study, we present preliminary data on the effect of automated 3D image alignment on the time to arrive at a decision about an imaging finding, the agreement of multiple of multiple observers, the prevalence of comparison examinations, and technical success rates for the image alignment algorithm. We found that automated image alignment reduced the average time to make a decision by 25% for cases where the structures are rigid, and when the scanning protocol is similar. For cases where these are not true, there is little or no benefit. In our practice, 54% of cases had prior examinations that could be automatically aligned. The overall benefit seen in our department for highly similar exams might be 20% for neuro and 10% for body; the benefit seen in other practices is likely to vary based on scanning practices and prevalence of prior examinations.

Appendix A

The Registration Algorithm Parameters

To improve speed and avoid local minima, registration algorithms often begin with low resolution images to get an approximately correct answer, and then use that to move on to a higher resolution image to improve that estimate until one finally uses the original high resolution image for the final estimate of the relationship between the two data sets. In our case, because the X and Y dimensions were much smaller, we “shrank” the image eight times more in the X and Y directions than in the Z direction.

There are many options available when performing a rigid alignment of one data set with another. One decision is how many types of motion you will allow to fit one data set onto the other. For instance, one could allow only translation along the three axis (X, Y, and Z). If one also allowed rotation about the three axes, there would be a total of six degrees of freedom. We included these six, plus a single scaling factor for a total of seven degrees of freedom.

Once the geometric relationship between two data sets is determined, it is necessary to “recut” one data set to map onto the other. To do this, it is necessary to create some formula for creating the values between the grid values in the actual image to create the new aligned image. One could simply take the voxel closest to the computed location (known as nearest-neighbor interpolation) but the result is often a jagged appearing output. B-spline is a mathematical technique for creating intermediate values between known values that appears more pleasing.