Abstract
Traditional sampling schemes often assume that the sampling locations are known. Motivated by the recent bioimaging technique known as cryogenic electron microscopy (cryoEM), we consider the problem of reconstructing an unknown 3D structure from samples of its 2D tomographic projections at unknown angles. We focus on 3D convex bilevel polyhedra and 3D point sources and show that the exact estimation of these 3D structures and of the projection angles can be achieved up to an orthogonal transformation. Moreover, we are able to show that the minimum number of projections needed to achieve perfect reconstruction is independent of the complexity of the signal model. By using the divergence theorem, we are able to retrieve the projected vertices of the polyhedron from the sampled tomographic projections, and then we show how to retrieve the 3D object and the projection angles from this information. The proof of our theorem is constructive and leads to a robust reconstruction algorithm, which we validate under various conditions. Finally, we apply aspects of the proposed framework to calibration of X-ray computed tomography (CT) data.
Introduction
In traditional computed tomography (CT), the acquisition geometry—specifically the angle of each X-ray projection—is precisely controlled and known. However, emerging modalities such as cryogenic electron microscopy (cryoEM) present a significantly more difficult challenge. In cryoEM, researchers capture thousands of 2D snapshots of biological molecules suspended in a thin layer of ice. These molecules are frozen in random, unknown orientations. Reconstructing the 3D structure from these "blind" projections is a fundamental computational problem in modern bioimaging.
Current reconstruction methods often rely on iterative optimization or discrete approximations that lack strong theoretical guarantees. This paper addresses the core sampling question: Under what conditions is perfect 3D reconstruction possible when both the signal structure and the projection angles are unknown? By focusing on signals with a Finite Rate of Innovation (FRI)—specifically 3D convex polyhedra and point sources—the authors provide a framework for exact reconstruction from a minimal number of projections.
Methods
The proposed approach leverages the sparse nature of FRI signals, which are defined by a finite number of degrees of freedom (such as the coordinates of vertices). The reconstruction follows a three-step constructive process:
- 2D Feature Extraction: The authors utilize the divergence theorem to convert the 2D area integral of a tomographic projection into a sum over its projected vertices. By calculating "complex moments" of the projection, they apply the annihilating filter method (Prony’s method) to extract the exact 2D coordinates of the vertices from sampled, potentially blurred images.
- Vertex Pairing: A critical challenge in blind reconstruction is determining the correspondence between features in different images. The authors developed a rank-deficiency criterion to "pair" vertices across projections, identifying which 2D point in one projection corresponds to the same physical 3D vertex in another.
- Joint Factorization: Once vertices are paired, the problem is formulated as a joint factorization task. Using Singular Value Decomposition (SVD), the algorithm simultaneously solves for the 3D coordinates of the vertices and the rotation matrices representing the unknown projection angles.
To handle noise, the authors introduced a robust variant of the algorithm that uses rational functions (poles) to approximate analytic functions, enhancing stability in non-ideal conditions.
Key Insights
A primary theoretical contribution is the "Power of Three": the authors prove that as few as three projections are sufficient to uniquely determine the 3D structure and the unknown angles (up to an orthogonal transformation). Remarkably, this minimum requirement is independent of the signal's complexity. Whether a polyhedron has 4 vertices or 100, three distinct views provide enough information for reconstruction. This contrasts sharply with bandlimited signals, where the required number of samples scales with the signal's bandwidth or complexity.
Key Results
The framework was validated using both synthetic simulations and real-world experimental data.
- Perfect Reconstruction: In noiseless scenarios, the algorithm achieved perfect reconstruction (zero error) for vertex locations and angles with $J \ge 3$ projections.
- Noise Resilience: The algorithm demonstrated significant robustness to noise. At a Signal-to-Noise Ratio (SNR) of 5dB, the vertex reconstruction error remained as low as approximately 7.6%.
- Real-World Application: The method was applied to a real X-ray CT dataset of a wooden block containing metal ball markers. The algorithm successfully estimated the projection angles with an average error of only 0.005 radians, demonstrating its potential for calibrating misaligned CT hardware.
Discussion and Limitations
While the results are promising, several constraints remain. The current mathematical proof is limited to convex polyhedra. Non-convex shapes present "self-shadowing" issues, where certain vertices may be hidden in specific projections, complicating the feature extraction step. Additionally, the vertex pairing step involves checking permutations ($K!$), which can become computationally prohibitive for objects with a very high number of vertices. Finally, the current model assumes parallel-beam geometry, which is standard in cryoEM but would need adaptation for the fan-beam or cone-beam geometries common in medical CT. Future work will focus on relaxing these constraints to broaden the applicability of the FRI-based reconstruction framework.
Summary updated