Surgical Navigation: Registration and Target Error

Phantom head, fiducial markers and a tracker: why FRE is not TRE.

Switch to 2D

Presets

Playback

Fiducial markers

How many markers, where they sit and how well each is localised.

Target

A point inside the brain-sized region that the registration is meant to find.

Simulation and drawing

How fast new registrations arrive and how much the errors are enlarged in the picture.

Camera

Quality

Layers

Model checks

The science

The problem

A navigation system holds an image of the patient (here a phantom head, in image coordinates) and measures positions in the room with a tracker (physical coordinates). Fiducial markers whose positions are known in both frames let the software find the rigid transform q = R p + t that carries image points into the room. This page is a geometry and statistics lesson on a phantom; it is not medical advice and not a description of any procedure.

Finding R and t (Horn / Kabsch)

Subtract the centroids p̄, q̄, form the 3 x 3 matrix H = sum (p_k - p̄)(q_k - q̄)^T and take its singular value decomposition H = U S V^T. The least-squares rotation is R = V diag(1, 1, d) U^T with d = sign det(V U^T), which prevents a mirror image, and t = q̄ - R p̄. The page also solves the problem with Horn's unit quaternion (the largest eigenvector of a 4 x 4 symmetric matrix) and checks that both methods give the same rotation.

Three different errors

FLE, the fiducial localisation error, is how far each measured marker is from its true place; here every marker gets independent isotropic Gaussian noise whose 3D RMS is the FLE slider. FRE, the fiducial registration error, is the RMS gap between the markers and their registered partners after the fit; the system can show this number. TRE, the target registration error, is the gap at the target point, which is what really matters and is never measured directly.

What theory predicts

For small noise the mean squared FRE is <FRE^2> = (1 - 2/N) FLE^2, so FRE shrinks as markers are added and says almost nothing about the target. The target error follows Fitzpatrick, West and Maurer (1998): <TRE^2> = (FLE^2 / N) (1 + (1/3) sum_k d_k^2 / f_k^2), where d_k is the target's distance from the k-th principal axis of the marker cloud and f_k the RMS distance of the markers from that axis. It is smallest at the marker centroid (FLE / sqrt N) and grows as the target moves away from the markers or the markers are nearly in a line or plane. The FRE-versus-TRE graph shows the practical lesson: the correlation between them is weak.

How the page checks itself

(1) With zero noise 25 random poses are recovered to rounding error. (2) The rotations from noisy data are orthogonal with determinant 1 and agree between SVD and Horn. (3) A Monte Carlo of 12 000 registrations is compared with the TRE formula at the target and at the centroid. (4) The same trials are compared with the FRE formula. The 2D page also compares its colour map with a Monte Carlo at five points of the plane.

Simplifications, honestly

The head is an ellipsoid with markers exactly on its surface; the target is kept inside a brain-sized ellipsoid; FLE is isotropic, equal for every marker and applied only to the tracker-space copy; the rigid body does not deform (no brain shift), and tracker lag, calibration error and marker-pad motion are ignored. The TRE formula is first order: it is rough when FLE is not small compared with the thinnest spread of the markers (the readout warns). Typical published figures, quoted from memory and to be checked: skin markers have FLE of the order of 1 to 3 mm, bone-anchored markers well below 1 mm. The sizes of the head and the bench scene are schematic; the tracker is drawn much closer than in reality.

What to try
  1. Start with the typical preset and watch the yellow cloud: each dot is where the registration puts the target in one simulated run; its spread is the TRE.
  2. Press 'FRE is small but TRE is large': five markers on the forehead give a tiny FRE, yet the target at the back of the head moves several times more than the markers do.
  3. Move the target towards the fiducial centroid (cyan cross in 2D) and the error falls to FLE / sqrt N; move it far away and it grows.
  4. Add markers one at a time: the curve in the first graph falls roughly as 1 / sqrt N.
  5. Switch to the midline arc: the markers are nearly in one plane, so targets far to the side are poorly constrained.
  6. Press the experiment button: 5 000 fresh registrations give a measured RMS TRE to compare with the formula.
Loading Blender model...
RMS TRE against number of fiducials
FRE against TRE of each registration
RMS TRE against distance from the fiducial centroid
Running RMS FRE and TRE
Made with 🟠 Blender · Three.js