In an ideal world, reconstructing a 3D object from 2D images would involve rotating the sample a full 360 degrees, capturing every possible perspective. In reality, physical constraints—such as the thickness of a sample holder, the flat geometry of a specimen, or the risk of destroying a biological cell with too much radiation—prevent this.
When a microscope can only tilt a sample between, say, -70 and +70 degrees, it creates a "Missing Wedge" of data. This isn't just a minor gap; it is a fundamental mathematical void in Fourier space that distorts the final 3D reconstruction, turning spheres into ellipsoids and blurring fine details into streaks.
The Geometry of Incompleteness
To understand why the missing wedge causes specific artifacts, we must look at the relationship between imaging angles and spatial frequencies. According to the projection-slice theorem, each 2D projection of an object provides a central slice of that object’s 3D Fourier transform. If you cannot tilt the sample to a specific angle, you cannot sample the spatial frequencies corresponding to that orientation.
As Cuicui Zhao and colleagues explain in their review of cryo-electron tomography (cryo-ET), this gap leads to anisotropic resolution. Because the "vertical" information (along the beam axis) is under-sampled, objects appear elongated. A spherical virus particle, for instance, might look like an American football. Andreas Velten and Ramesh Raskar observed a similar "missing cone" effect in non-line-of-sight imaging, where restricted observation planes prevent the capture of features parallel to the scanning surface.
Filling the Gap with Geometric Priors
If the data is missing, how can we reconstruct the truth? The earliest successful strategies relied on geometric priors—assumptions about what the object should look like.
If you know you are looking at a collection of nanoparticles, you can assume the objects are discrete and perhaps even spherical. Sara Bals demonstrated that by incorporating Sparse Sphere Reconstruction (SSR), researchers could overcome missing wedge artifacts using ten times fewer projections than standard methods. Similarly, Andreas Alpers showed that assuming convexity or homogeneity allows algorithms to "hallucinate" the missing data points with high mathematical confidence.
Deep Learning and Neural Hallucination
In the last decade, the field has shifted from hard-coded geometric rules to learned priors. Instead of telling the computer that an object is "convex," we show it thousands of examples of what biological machinery looks like.
Emmanuel Moebel and Antonio Martínez-Sánchez pioneered the use of 3D convolutional neural networks (CNNs) to identify macromolecules despite the elongation caused by the missing wedge. These models learn to recognize the "distorted" version of a protein as being identical to its "clean" version.
More recently, researchers have moved toward self-supervised and generative approaches that don't require massive labeled datasets:
- DeepDeWedge: Reinhard Heckel and Simon Wiedemann developed DeepDeWedge, which fits a neural network directly to noisy 2D projections. By using the physics of the imaging process as a constraint, the network learns to fill in the missing wedge while simultaneously denoising the image.
- Foundation Models: Quanquan Gu and Yi Zhou introduced CryoFM, a flow-based generative model. It learns the underlying distribution of all known protein structures, allowing it to act as a "plug-and-play" prior that restores missing data by finding the most likely biological structure that fits the observed projections.
- Diffusion Models: Zhantao Deng proposed TEMDiff, which uses 3D diffusion to iteratively refine a reconstruction, effectively "sampling" from the space of possible 3D volumes to find one that satisfies the limited-angle projections.
Rethinking the Voxel
A final, emerging direction moves away from traditional grid-based (voxel) reconstructions entirely. Nalini M. Singh has adapted Gaussian Splatting to electron tomography. By representing atoms as mathematical Gaussians rather than points on a grid, the reconstruction becomes more resilient to the "streak" artifacts typical of the missing wedge, as the optimization process naturally favors compact, physically plausible atomic shapes over the elongated "ghost" structures that plague voxel-based methods.
By combining the rigid physics of diffraction—as seen in the work of Jianwei Miao on vector tomography—with the flexible priors of generative AI, researchers are finally closing the gap, turning the "missing wedge" from a brick wall into a solvable inference problem.