by Reuben R Shamir.
I was two years into my PhD studies at Jerusalem working hard but with few results on the intended topic of my dissertation, estimation of target registration error (TRE). As opposed to other studies that estimated average TRE, I was interested in a patient-specific approach and searched for methods to upper bound the setup-specific TRE based on the fiducial point configuration. I never found such a function, but fortunately my supervisor, Leo Joskowicz, and I proved that no such function exists, so I became a doctor after all and could leave studies of registration error behind me.
At the same time, other graduate students in our lab did some nice work on image-to-image registration and took first place in the “The Retrospective Image Registration Evaluation Project” (RIRE). But the development of their algorithm took about a year and required hard work.
My struggles with TRE estimation, led to the discovery of a method to beat the RIRE system with very little effort and some basic programming. The key idea is that TRE reported by RIRE is associated with the distance of target locations in the aligned images. By applying one arbitrary transformation followed by three orthogonal transformations on one image and recording TRE after each transformation, it is possible to recover the exact target locations.
The near zero TRE values on the diagonal in the screenshot below from the RIRE website show that the computed target locations are practically the same as the actual ones (follow this link to the actual result online).
These actual target locations were never disclosed to my supervisor or any other lab member, as this would be inappropriate. I also never used it to get my name to the top of the list, but now you can do it.
Reuben Shamir, PhD is a Postdoctoral Fellow at the Neuromodulation Center at the Biomedical Engineering Department at Case Western Reserve University, Cleveland, OH, USA.