Gold-fiducial alignment
Before reconstruction every tilt image must land in one common 3D frame — using gold beads as fiducials, or markerless patch tracking
Once the tilt series is collected you have a stack of 2D images, but they are not registered to each other: the stage has mechanical backlash between tilts, and beam-induced motion makes the specimen drift and deform during exposure. Each image carries its own unknown in-plane translation, rotation, and magnification error. Feed that stack straight into backprojection and rays from different tilts will not intersect at the same point — the whole volume smears. Alignment is the step, before reconstruction, that solves for each image’s geometric transform and maps every image into one common 3D coordinate frame. This page is about how it is actually done; the geometric model behind it is in tilt-series alignment.
In the pipeline this comes right after data processing (CTF and dose weighting): you already have a stack of motion-corrected, CTF-estimated tilt images. In comes that stack, out comes a set of per-image geometry parameters, and the standard tool for the job is IMOD’s etomo (or AreTomo for the markerless case).
Picture a few fixed bright points embedded in the specimen. With perfect tilt geometry, each point’s projection traces a predictable curve as the stage tilts. The deviation of the real curve from the ideal one is a direct measurement of how far that frame’s field of view has drifted. Combine several points and you can invert the problem: solve for how each image should be shifted and rotated. That is the whole idea of fiducial alignment.
Why everything must land in one frame
Backprojection only works if you know which direction each image was shot from and which point in space it was aimed at — that is what lets you smear each image back along its ray into the volume. If image 17 is actually a few pixels off from where the model thinks it is, its ray lands in the wrong place; rays from other tilts that should have met at some feature now miss each other. The feature gets painted into a blur. Every image carries a little of this error, and stacked up the whole volume smears. So alignment’s job is to find, for every image, a transform that makes all the tilt rays point back at the same set of points in one shared 3D frame.
Gold beads as fiducials
The most reliable approach is to mix colloidal gold beads, ~5–10 nm across, into the suspension before vitrification. Gold scatters electrons far more strongly than biological material, so each bead appears as a sharp dark dot in every projection — high-contrast, point-like, identifiable at every tilt. These are the fiducial markers that run through the entire series: because they are the same fixed physical points, their projected positions across tilts tie the images together.
Bead-based alignment (the canonical implementation is IMOD’s etomo workflow) runs in this order:
- Coarse alignment. Ignore the beads for now and just cross-correlate adjacent tilts pairwise to remove the large obvious shifts, so the stack is roughly stacked up. This only gives the tracking step a clean starting point — beads will not jump far between frames.
- Seed & track. Detect the beads automatically in the zero-tilt (sharpest) image and pick a well-distributed, non-overlapping set away from the field edges as seeds; then follow each bead through the whole tilt series, recording its measured projected coordinate at every tilt. Tracking tends to break at high tilt — beads lose contrast and crowd together — and those gaps usually need to be repaired or pruned by hand back in 3dmod.
- Solve. Fit each image’s geometry jointly from all bead tracks. The objective is to minimize reprojection error: given a set of 3D bead coordinates and a per-tilt transform, projecting the 3D beads forward through those transforms should land on the measured positions. The optimizer adjusts both the 3D bead model and all the image transforms until the two agree as closely as possible.
- Inspect the residual. The solve reports a residual number (next section). Read the overall residual and the per-bead residuals, drop the beads that tracked wrong, and re-solve until the residual settles.
After those four steps etomo holds all the geometry it needs to place each image back into the volume.
What the solve recovers
Each image gets more than a pair of translations. The full model assigns every projection:
- in-plane translations in and , correcting the field shift;
- an in-plane rotation;
- a magnification term, absorbing the small scale change from focus/height variation through the tilt.
All of these are referenced to a single global tilt-axis angle (the orientation of the rotation axis in the detector plane), and the tilt angles themselves are refined too — the stage readout rarely matches the true geometry exactly. The key is that these parameters are solved jointly, not by registering neighboring images pairwise. Pairwise errors accumulate along the series into a systematic distortion; a single global fit spreads the error out and lets the beads constrain each other.
The number that matters is the residual: the difference between each measured bead position and where the fitted model reprojects it, reported as a root-mean-square over all beads and all tilts (in pixels or nanometers). The residual is a direct readout of alignment quality.
What counts as good or bad is visible qualitatively. A good alignment: the overall residual is sub-pixel (for a typical collection, roughly a few tenths of a pixel up to a little over one pixel), every bead’s residual is close to the others with no clear outlier, and the residual-vector scatter looks like a random little cloud with no preferred direction. A bad alignment: the overall residual is several pixels; one or two beads have residuals far above the rest (tracking jumped to a neighbor or a noise speck at some tilt); or the residual vectors all point the same way and vary systematically with tilt — that says the error is not random noise but a misfit geometry (wrong tilt-axis angle, or a parameter that should have been enabled is missing). In practice, go after the outlier beads first: drop them and re-solve, and the overall residual often falls immediately; if the pattern is a systematic direction instead, go back and check the tilt axis and the solve options.
Why the residual sets the resolution: reconstruction places each image back into the volume according to its fitted geometry. If a feature’s true position is off from the model by even a fraction of the high-resolution sampling step, rays from different tilts no longer coincide and the feature blurs. The highest spatial frequencies go first — they are exactly where the rays must overlap most precisely. So the residual is a hard ceiling on everything downstream; no amount of subtomogram averaging recovers high frequencies lost at the alignment stage.
Markerless alignment
Adding beads is not always possible: in-situ lamellae cannot be doped with gold, and beads may land right on the region of interest. Fiducial-less alignment instead registers images using the specimen’s own content.
The most common scheme is patch tracking: cut each projection into many small patches and cross-correlate each patch between adjacent tilts, using image features in place of beads to build up the same per-tilt transforms. Projection-matching variants iterate against a provisional reconstruction. Tools like AreTomo follow this route, are often fast enough to align during acquisition, and can correct local deformation along the way. The cost: without clean high-contrast anchors, markerless alignment is less stable on very thin, low-contrast, or featureless specimens, and its quality depends on the content being textured enough to correlate. A practical rule: if your specimen already has gold beads, use them — bead alignment is almost always more accurate; fall back to markerless only when you cannot add beads or the beads obscure the target.
Handing off to reconstruction
The output of alignment is a set of refined geometry parameters — per-image transforms, tilt-axis orientation, corrected tilt angles. These feed directly into weighted backprojection (WBP) or iterative reconstruction (the tilt step in etomo), stacking the 2D images into a 3D tomographic volume. Note that alignment does not fill the missing wedge: it only registers the angles you did collect; the Fourier void from the never-sampled tilts is still empty and must be inferred by learned priors like CryoGEN / CryoWGEN. In other words, alignment sets how faithfully you can use the data you have, while the missing wedge sets how much that data is missing in the first place.
Previous: data processing — CTF and dose weighting · Next: from tilt series to tomogram