Abstract
It is well known that a band-limited signal can be reconstructed from its uniformly spaced samples if the sampling rate is sufficiently high. More recently, it has been proved that one can reconstruct a 1D band-limited signal even if the exact sample locations are unknown, but given just the distribution of the sample locations and their ordering in 1D. In this work, we extend the analytical bounds on the reconstruction error in such scenarios for quasi-bandlimited signals. We also prove that the method for such a reconstruction is resilient to a certain proportion of errors in the specification of the sample location ordering. We then express the problem of tomographic reconstruction of 2D images from 1D Radon projections under unknown angles with known angle distribution, as a special case for reconstruction of quasi-bandlimited signals from samples at unknown locations with known distribution. Building upon our theoretical background, we present asymptotic bounds for 2D quasi-bandlimited image reconstruction from 1D Radon projections in the unknown angles setting, which commonly occurs in cryo-electron microscopy (cryo-EM). To the best of our knowledge, this is the first piece of work to perform such an analysis for 2D cryo-EM, even though the associated reconstruction algorithms have been known for a long time.
Introduction
In the field of Cryo-Electron Microscopy (cryo-EM), researchers face the challenge of reconstructing the structure of molecules from thousands of projections captured at unknown orientations. This problem, known as Unknown View Tomography (UVT), differs significantly from standard medical CT scans where projection angles are precisely controlled. While various heuristic algorithms have been developed to tackle UVT, a rigorous mathematical understanding of its limits—such as the required number of projections and the impact of noise—has remained elusive.
This paper provides the first formal theoretical error bounds for 2D UVT. By establishing a link between 2D tomography and 1D signal reconstruction from unknown sample locations, we demonstrate that accurate reconstruction is mathematically guaranteed provided the statistical distribution of the angles is known.
Theoretical Framework
1D Foundation and Quasi-Bandlimited Signals
The research begins by analyzing the reconstruction of 1D signals from samples at unknown locations. If the probability density function (PDF) of the sampling locations and their relative order are known, order statistics can be used to estimate the most likely location of each sample. A key contribution of this work is the extension of this theory to quasi-bandlimited (QBL) signals. Unlike strictly bandlimited signals, QBL signals have frequency content that decays rapidly but does not vanish entirely, making them a far more realistic model for physical images with sharp edges and fine textures.
The 2D UVT Bridge
To move from 1D theory to 2D images, the authors utilize the Fourier Slice Theorem. This theorem states that the 1D Fourier transform of a projection at a specific angle corresponds to a slice through the 2D Fourier transform of the original image. Consequently, the 2D UVT problem can be modeled as a collection of 1D periodic signal reconstruction problems along concentric rings in the Fourier domain.
The Reconstruction Pipeline
The proposed Reconstruction Algorithm (RA) consists of four primary stages:
- Denoising: PCA-based techniques are applied to the noisy 1D projections to improve signal quality.
- Ordering: The algorithm uses Laplacian Eigenmaps (LE) to determine the relative angular order of the projections. This step identifies which projections are "neighbors" based on their structural similarity.
- Angle Assignment: Using the known distribution (e.g., Uniform or von Mises), the algorithm assigns the most probable angles to the ordered projections.
- Reconstruction: Finally, a Non-Uniform Fast Fourier Transform (NUFFT) is used to map the estimated Fourier samples back into a spatial 2D image.
Key Results and Analysis
Convergence and Noise Resilience
The study establishes that the expected squared reconstruction error $E$ decays at an asymptotic rate of approximately $O(\log^3 N / N)$, where $N$ is the number of projections. This implies that doubling the number of available projections roughly halves the reconstruction error. Furthermore, the error scales linearly with the noise variance ($\sigma^2$), demonstrating the method's stability under typical experimental conditions.
Empirical Validation
The theoretical bounds were validated using a 608x608 pixel ribosome image from the Electron Microscopy Data Bank (EMDB). Simulations using both uniform and non-uniform (von Mises) angle distributions confirmed that the empirical error closely tracks the theoretical predictions. Notably, the Laplacian Eigenmaps algorithm proved highly resilient; even with 10% noise, it recovered a near-perfect ordering of projections, leading to high-quality image reconstruction.
Conclusion and Future Work
This work fills a critical gap in the theory of tomographic reconstruction by providing the first analytical bounds for the 2D unknown angle problem. It proves that knowing the shape of the randomness (the distribution) is a powerful substitute for knowing exact angles.
While this provides a solid foundation for 2D cryo-EM, extending these results to 3D remains a significant challenge. In 3D, rotations involve more complex degrees of freedom than simple 2D angles, and the sorting math becomes considerably more intricate. Future research will also explore scenarios where the angle distribution itself is unknown or biased, as well as developing more robust ordering techniques for extremely high-noise environments.
Summary updated