Wedge-Aware Planar Splatting for Limited-Angle Tomography of Moving Specimens

Many 3D imaging systems cannot view an object from every direction. A cryo-electron microscope can tilt its sample only to about ±60–70°, and a camera on a robot or underwater vehicle often sees a scene from a narrow range of viewpoints. The directions that are never seen leave a "missing wedge": a cone of 3D spatial frequencies (fine detail along particular directions) that no measurement ever records. Reconstructions then stretch along the viewing axis and lose surfaces that face the beam. In cryo-electron tomography, the tops and bottoms of cell membranes simply disappear. This note proposes rebuilding such volumes from a few thousand explicit, flat, oriented primitives, "planar splats" (small disks that are rendered by projecting them directly onto the image, not by querying a neural network). Each primitive carries its own estimate of how much of it the data can actually see. Primitives the measurements pin down are left free. Primitives that sit mostly in the missing wedge are pushed toward being flat sheets and to continue the orientation of their well-observed neighbors, so a membrane seen at oblique tilts is carried across the region the microscope cannot see.

The second ingredient addresses an assumption that every limited-angle method in our file makes and that is false in practice: that the specimen stays still across the tilt series. A tilt series is recorded one angle at a time, so beam-induced drift or slow deformation shows up as inconsistency between angles, and a static solver absorbs it as blur or streaks. Because the representation is explicit and small, we can give the primitives a low-dimensional motion as a function of acquisition time and still reconstruct in minutes rather than hours. The claim is specific. Weighting a planar prior by per-primitive unobservability, with motion across the tilt series modeled jointly, recovers beam-facing surfaces that voxel, coordinate-network, and isotropic-Gaussian solvers lose. Generic planarity regularization does not account for this gain, and none of it costs more runtime than the fastest unsupervised solver.

What exists, and where this differs

Three lines of limited-angle work motivate and constrain the design. Missing Wedge Completion via Unsupervised Learning with Coordinate Networks fits a coordinate neural network (a network mapping $(x,y,z)$ to density) directly against ±60° projections. It avoids the pretraining bias of supervised inpainting and cuts P22 phage reconstruction from over 8 hours (IsoNet) to under 20 minutes, but it oversmooths sharp boundaries and assumes a static specimen. A Gaussian Parameterization for Direct Atomic Structure Identification in Electron Tomography shows that explicit Gaussians plus a hard isotropy constraint reject streak artifacts at ±70°. That paper supplies the method template: a differentiable projector and explicit primitives. It also sets the constraint: spherical primitives cannot represent membranes, sheets, or interfaces. Z-Splat diagnoses the same "missing cone" in camera-only Gaussian splatting via the Fourier slice theorem and fixes it by adding sonar measurements. It uses volumetric ellipsoids, names dynamic scenes as future work, and requires a second sensor, which a vacuum microscope column cannot have.

The planar-primitive literature supplies the shape machinery. 2D Gaussian Splatting and PGSR define disks with explicit normals and perspective-correct ray–disk intersection. Both note that they degrade under limited viewpoint coverage, which is exactly our regime. Effective Rank Analysis and Regularization gives a differentiable "effective rank" of each primitive's shape (about 1 for needles, 2 for disks, 3 for blobs) and shows that pushing away from needles helps geometry on DTU, reducing Chamfer distance from 1.96 to 1.04. We reuse this measure, but we apply it with a weight set by the acquisition geometry rather than uniformly.

For motion, Split4D attaches linear velocities to Gaussians, and AT-GS warps surfels frame-to-frame with a small hash-grid network. Both assume dense camera domes that see every frame from many directions. In a tilt series, each moment in time is seen from exactly one direction, so motion must be far more constrained. We borrow the per-primitive velocity idea and restrict it to a smooth, low-order displacement field over tilt index.

No paper in the file places planar primitives under a physical missing wedge, and none reconstructs a time-varying specimen from a limited-angle tilt series. The closest work, the atomic Gaussian parameterization, solves the wedge for point-like atoms by forcing isotropy. We handle continuous surfaces with an orientation-aware, observability-weighted rank prior and add motion.

Core Mechanism

A Gaussian primitive with spatial covariance $\Sigma_i$ has a Fourier power spectrum that is itself a Gaussian with covariance proportional to $\Sigma_i^{-1}$. This is a standard property of the Gaussian, and it lets us ask, in closed form up to a cheap angular integral, how much of primitive $i$'s energy falls where the tilt range actually samples. Let $S$ be the set of sampled frequencies, the union of central slices perpendicular to each beam direction. Define the observed fraction

$$\rho_i = \frac{\int_S |\hat g_i(k)|^2\,dk}{\int |\hat g_i(k)|^2\,dk},$$

where $\hat g_i$ is the Fourier transform of primitive $i$. A flat sheet whose normal points along the beam concentrates its spectrum along the beam axis, inside the wedge, so it has $\rho_i$ near zero. A vertical sheet or a small sphere has $\rho_i$ near one. Stretching a blob along the beam axis raises its $\rho_i$ while barely changing its projections. This is exactly why unconstrained splats and voxels elongate. The data term is indifferent, and nothing else stops them.

The reconstruction minimizes

$$\mathcal{L} = \sum_{\theta} \big\| P_\theta\big(\mathcal{G}(t_\theta)\big) - y_\theta \big\|^2 + \lambda \sum_i (1-\rho_i)\, R_{\text{rank}}(\Sigma_i) + \mu\, R_{\text{coplanar}} + \nu\, R_{\text{motion}}.$$

Here $y_\theta$ is the measured projection at tilt $\theta$, recorded at time $t_\theta$. $P_\theta$ is a differentiable splatting projector, and $\mathcal{G}(t)$ is the primitive set displaced to time $t$. $R_{\text{rank}}$ penalizes effective rank below 2, which forbids needles. $R_{\text{coplanar}}$ ties each low-$\rho$ primitive's normal and plane offset to those of its nearest well-observed neighbors, which carries a membrane across its unseen cap. $R_{\text{motion}}$ penalizes roughness of a displacement field parameterized by a few control points with low-order polynomials in $t$. The weight $(1-\rho_i)$ is the part that is new. It lets atoms and well-seen structure keep whatever shape the data supports, including rank 3, and applies the planar prior only where the data are silent. The coplanarity term is what turns the prior from "look flat" into "continue the surface." A generic rank penalty cannot do that on its own.

Research Program

The first step is the projector. Tomographic projections are line integrals, and a Gaussian's line integral is a 2D Gaussian with a closed form, so we adapt the tile-based splatting rasterizer to parallel-beam geometry with a per-tilt affine alignment. We also build a perspective variant for the optical cross-check.

The second step implements $\rho_i$ for a double-wedge sampling set, evaluated by a fixed quadrature over directions and recomputed every few hundred iterations, because shapes change slowly. We also freeze densification in the first phase, as RobustSplat does. That paper showed that inconsistent observations make densification spawn stranded primitives. Inconsistency between tilts caused by motion is the tomographic analogue of transient distractors, so we fit coarse structure and motion first and densify only afterward.

The third step is an identifiability analysis of motion. With one view per time step, a smooth rigid drift is recoverable, but deformation along the unseen axis at a given tilt is not. We will characterize, on simulations, which motion amplitudes and spatial frequencies the joint model can separate from structure. We will report where it fails rather than claiming general 4D tomography.

Baselines

The naive reference is weighted back-projection and SART (simultaneous algebraic reconstruction, the standard iterative voxel solver) on the same tilt series. This fixes the size of the missing-wedge artifact every other method must reduce.

The classical regularized competitor is SART with total variation (a penalty on image gradients that favors piecewise-constant volumes), with its weight tuned on a held-out simulated volume, as in Geometric reconstruction methods for electron tomography.

The strongest existing unsupervised method is the coordinate network of Missing Wedge Completion via Unsupervised Learning with Coordinate Networks. It is the most relevant competitor because it is also pretraining-free and optimized against raw projections. For this baseline, the only figure in our file is runtime: under 20 minutes on P22 versus over 8 hours for IsoNet. No reported PSNR, SSIM, or FSC value was found on Lacuna for it, so we will rerun it on identical simulated data with its published architecture and a matched number of tuning trials. Beating it means equal or better directional FSC along the beam axis, higher membrane-cap recovery, and no more wall-clock time on one GPU.

The explicit-primitive competitors are two. The first is the isotropic Gaussian solver of A Gaussian Parameterization, rerun with our projector so that only the shape constraint differs. The second is unconstrained Gaussian splatting with our projector.

The key ablation is uniform effective-rank regularization in the style of Effective Rank Analysis and Regularization, with the weight tuned to its best value. It shows whether wedge-aware weighting and coplanar continuation add anything beyond generic flattening. Further ablations drop $R_{\text{coplanar}}$ and drop the motion model on drifting data. All methods receive the same tilt series, the same alignment, one GPU, and the same hyperparameter search budget.

Why not just…?

Why not just use total-variation regularization on a voxel grid? Total variation favors flat density plateaus, not thin sheets. A membrane cap that the data never record looks, to total variation, like a gradient worth removing, so the regularizer actively erases it. The atomic Gaussian paper reports that voxel solvers like SART suffer streak-induced false positives at ±70° that explicit primitives suppress. Our hollow-sphere experiment measures cap recovery directly. If tuned total variation recovers caps comparably, the geometric argument here is wrong and we will say so.

Why not just improve the coordinate network? It is the most serious threat on the quality axis. Its known weakness, oversmoothing of sharp boundaries, is reported by its own authors, and a smooth network has no explicit notion of a surface to continue. A coordinate network with a Hessian or surface prior is a fair variant, and we will include one. Explicit splatting should also be cheaper per iteration than querying a network, which the runtime comparison checks.

Why not just apply generic effective-rank regularization to tomographic splats? This is the alternative most likely to work nearly as well. The uniform-regularization ablation exists to settle it. We expect uniform weighting to fail in two ways: it will flatten genuinely compact structure (atoms, dense granules) that should stay rank 3, and it will not fill caps, because flatness alone does not choose an orientation. The atomic nanoparticle data test the first failure and the hollow spheres test the second. If uniform regularization matches us on both, the contribution reduces to "rank regularization transfers to tomography," and we would report that plainly.

Why not just penalize primitive extent along the beam axis? It is simpler, but it would equally squash real vertical membranes, which are well observed. The $\rho_i$ weighting reduces to this penalty only for primitives that are actually unobserved. The vertical-versus-horizontal membrane breakdown below makes the difference measurable.

Why not just align the tilt series better and keep a static model? Alignment removes a single rigid shift per tilt. It cannot remove deformation that varies across the field of view. The drift experiments separate the two by injecting rigid and non-rigid motion independently.

Proposed Experiments

The primary benchmark is the simulated cryo-electron tomography suite of P22 phage, hollow spheres, and ubiquitin used by the coordinate-network paper, reconstructed at ±60° and additionally at ±45° and ±70°. Metrics are directional Fourier shell correlation (agreement with ground truth as a function of frequency direction), the elongation ratio of spherical objects, and runtime. For hollow spheres we add a surface metric: the fraction of ground-truth shell area recovered within one voxel, reported separately for the beam-facing cap and the equatorial band. The decisive check is that cap recovery improves substantially over every baseline, while equatorial accuracy and directional FSC outside the wedge do not degrade.

The second benchmark is the atomic nanoparticle data (FePt, Au, and related) at ±70° from the Gaussian parameterization paper. There we require that $\rho$-weighting leaves atoms near rank 3 and does not reduce atom-identification precision or recall relative to the isotropic solver. The file gives no numerical rates for that solver, so it will be rerun.

The third is dynamics. On the P22 and hollow-sphere volumes we inject rigid drift and smooth non-rigid deformation of controlled amplitude, ordered by tilt time, and compare static and motion-aware versions of every method. Success means the motion model recovers static-quality FSC up to a stated amplitude, and the identifiability analysis predicts where it breaks.

As an optical cross-check that the mechanism is not specific to microscopy, we restrict DTU camera views to a cone of elevations, creating a missing cone, and compare 2D Gaussian Splatting, PGSR, and our weighted variant on Chamfer distance. The only full-view reference in the file is 1.04 mean Chamfer with rank regularization. No restricted-view result was found on Lacuna for any method, so all restricted-view numbers will be produced here under one protocol.