Tilt-series acquisition
A tomogram is built from many 2D projections recorded as the specimen is tilted through a range of angles.
A tilt series is the raw data of electron tomography: a set of 2D projection images of one specimen region, each recorded at a different tilt of the microscope stage. Rotating the specimen relative to the beam exposes the object from many directions, and these views are later combined into a 3D volume during reconstruction.
The single fact that organizes everything below is that each image is a projection: the electron beam integrates the specimen’s scattering density along its path and collapses a 3D object onto a 2D plane. One projection cannot tell a feature near the top of the ice from one near the bottom — depth is gone. Tilting recovers depth the way two eyes recover it: each new angle is a fresh line of sight, and a feature’s apparent position shifts from view to view by an amount that depends on how deep it sits. The reconstruction is the inversion of that shift. So the design of a tilt series is really the design of which lines of sight you spend your limited dose on.
Think of photographing a frosted-glass paperweight with objects suspended inside it. A single photo flattens everything onto one plane and you cannot say what is in front of what. Walk a quarter-circle around it taking photos as you go, and the relative motion of the inner objects between shots tells you their depth — near objects swing across the frame, far ones barely move. A tilt series is exactly this walk-around, except the specimen rotates instead of the camera, and two hard constraints bite: you can only walk part of the way around (the missing wedge), and every photo you take damages the specimen a little, so you get only a few dozen and each is dim and grainy.
The dose-fractionation trade-off — vary the number of tilts N:
The specimen tolerates only a fixed total dose before radiation damage. Splitting it over N tilt images gives each image dose ∝ 1/N, so per-image noise grows ∝ √N. Too few tilts: each is clean but angular sampling is sparse (streaks); too many: sampling is dense but each image drowns in noise. The sweet spot is in between.
The achievable angular range is limited by geometry. As the stage tilts, the path the beam travels through the slab of ice lengthens by roughly , so at high tilt the effective specimen thickness grows until electrons can no longer pass. Here is the tilt angle measured from the untilted (flat) position and is the cosine of that angle; at the factor is , at it is , and at it is about . A 200 nm slab therefore presents an effective thickness of roughly 400 nm at and nearly 600 nm at — and beyond some point the beam scatters so much it carries no usable contrast. Stage hardware and the supporting grid impose further limits. In practice acquisition stops near ±60° to ±70°, in angular increments of 1–3°, giving on the order of 40–60 projections. The angles never reached form the missing wedge, the most consequential artifact of tomographic geometry.
Because the specimen tolerates only a limited total electron dose, the budget must be fractionated across the whole series. If the total tolerable dose is and images are taken, each tilt receives roughly , so every projection is extremely noisy and reconstructions begin from low signal-to-noise data. Here is the cumulative electron exposure — measured in electrons per square Ångström () — that the vitreous specimen can absorb before radiation damage destroys the high-resolution information you came to image, and is the number of tilts. The two knobs pull against each other: more tilts ( large) sample the angular range more finely and shrink the gaps the reconstruction must interpolate, but each image then gets a thinner slice of dose and is correspondingly noisier; fewer tilts make each image cleaner but leave coarser angular gaps. The demo above lets you feel this directly — there is no free lunch, only a budget to allocate.
This is the central reason tomographic data is hard. A typical single-particle exposure might use the full dose budget on one micrograph; here the same budget is sliced into 40–60 pieces, so an individual tilt image can sit near a signal-to-noise ratio of order one or below (see signal-to-noise). The structure you want is buried in noise of comparable amplitude in every single frame, and only becomes visible after many noisy views are combined and after the redundancy of subtomogram averaging is brought to bear.
The order in which tilts are collected matters because radiation damage accumulates. Dose-symmetric schemes acquire the low-tilt views first — where the specimen is thinnest and information most valuable — then alternate outward to higher tilts in a symmetric pattern. This delivers the early, damage-free dose to the angles that most constrain high-resolution structure, while spreading later, damage-bearing exposures to the periphery of the angular range.
To see why the order is not a detail, contrast it with the older alternatives. A naive unidirectional scheme sweeps from straight to ; by the time it reaches the far end the specimen has absorbed nearly the whole dose, so one entire side of the angular range is recorded on a radiation-damaged sample and contributes only blurred, low-resolution information. A bidirectional scheme starts at and goes to one extreme, then returns to and goes to the other — better, but it still introduces a discontinuity in accumulated dose across the central tilts, which complicates alignment. The dose-symmetric scheme (commonly written as a grouped pattern like ) keeps accumulated dose a smooth, symmetric function of tilt angle and front-loads the precious early electrons onto the thin, near-axial views that carry the highest-resolution signal. The price is more stage movements and settling time per series, which is why it became practical only once stage stability and automation improved.
Several effects must be tracked across the series. Each image is recorded at a controlled underfocus, or defocus, which sets the contrast transfer function (see the CTF section). The stage shifts and the specimen drifts and deforms between tilts, so the frames are never perfectly registered. Correcting these movements is the job of tilt-series alignment, which must precede reconstruction. The low signal-to-noise ratio inherent to dose-limited acquisition is also what motivates learned restoration approaches such as CryoGEN: when each frame is this noisy, the question of how to recover the underlying density becomes a statistical-inference problem, not just a filtering one.
Per-image CTF and defocus
A single tilt series has no single defocus. The microscope is set to a nominal underfocus, but tilting the stage means different parts of a high-tilt image sit at different heights along the beam, so the focus varies across the field. Each projection therefore carries its own contrast transfer function, and at high tilt that CTF varies within the image as well. Estimating defocus per tilt — and, for the most demanding work, per strip across a tilted image — is a step in its own right, because the CTF inverts contrast at spatial frequencies that depend on defocus and must be corrected before those frequencies can be combined coherently in reconstruction. The defocus is also deliberately staggered across a dataset so that the CTF zeros of one exposure are filled by another.
The geometry is worth picturing concretely. With the stage flat, the whole field sits at one nominal height and shares one defocus. Tilt it to and the field becomes a ramp: a point a horizontal distance from the tilt axis sits at a height offset of along the beam, so its defocus differs from the axis value by that same . At a feature 1 µm off-axis is already a few hundred nanometres out of the nominal focus — a defocus difference large enough to shift the CTF zeros into different places. This is exactly why high-tilt images must be CTF-modeled in strips parallel to the tilt axis: each strip is at a roughly constant height and so a roughly constant defocus, and treating the image as one uniform CTF would mis-correct everything away from the axis.
From dose fractionation to the reconstruction
The dose budget is split twice. It is fractionated across the tilts so each angle receives roughly , and within each tilt the exposure is itself recorded as a movie of several sub-frames on a direct-detector camera. Splitting the exposure into frames allows beam-induced motion — the bulk drift and local deformation that a freshly illuminated specimen undergoes as charge builds up and the support relaxes — to be tracked and corrected before the frames are summed into one projection. Typical pixel sizes fall in the few-Ångström range, set by the magnification needed to resolve the target against the available dose. The aligned, CTF-aware projections, each tagged with its accumulated dose so that radiation-damaged high frequencies can be down-weighted, are the input that reconstruction back-projects or iteratively fits into a 3D volume.
The per-frame dose tagging deserves a word, because it is how the pipeline reconciles two competing facts: low spatial frequencies (overall shape) survive radiation damage well, while high spatial frequencies (fine detail) are erased early in the exposure. Dose weighting uses each frame’s accumulated dose to apply a frequency-dependent, exposure-dependent filter — keeping the high frequencies only from the early, undamaged frames and trusting later frames only for the coarse, damage-resistant signal. Combined with the dose-symmetric ordering described above, this means the final tomogram draws its high-resolution information predominantly from the low-tilt, early-exposure views, exactly where the specimen was thinnest and freshest. The whole acquisition design — angular range, increment, tilt order, frame splitting, dose tagging — is a coordinated effort to spend a fixed dose budget so that the views carrying the most recoverable information are the ones recorded under the best conditions.
Everything on this page sits between two neighbours in the pipeline. Upstream, the specimen was vitrified and the microscope set up as described in cryo-EM and vitrification; the tilt series is what that frozen sample is turned into. Downstream, these noisy, individually-aligned projections are registered by alignment and then merged by reconstruction into the 3D volume — with the missing wedge, the dose noise, and the per-tilt CTF all stamped into that volume as the artifacts that later restoration and averaging must contend with.