Accurate real space iterative reconstruction (RESIRE) algorithm for tomography

Minh Phạm, Yakun Yuan, Arjun Rana, Stanley Osher, Jianwei MiaoScientific 2023 · DOI

Abstract

Tomography is a cornerstone of 3D imaging in physical and biological sciences, yet it remains challenged by limited projection data and noise. We present the Real-Space Iterative REconstruction (RESIRE) algorithm, which bridges the gap between the computational efficiency of real-space methods and the high accuracy of Fourier-based approaches. By employing mathematically precise forward and back-projection steps within a least-squares optimization framework, RESIRE achieves superior 3D reconstructions. Our results on simulated and experimental datasets, including biomineral coral and amorphous thin films, demonstrate that RESIRE outperforms existing algorithms like SIRT and GENFIRE in both accuracy and resolution, particularly when data is sparse and noisy.

Introduction

The fundamental goal of tomography is to reconstruct a 3D object from a set of 2D projections. In practice, this process is often hindered by the "missing wedge" problem, where geometric constraints or radiation dose limits prevent the acquisition of a full range of projections. This leads to blurring and artifacts in the final reconstruction. Existing methods generally fall into two categories: real-space methods (e.g., FBP, SIRT), which are fast but often rely on mathematical approximations, and Fourier-based methods (e.g., GENFIRE), which are accurate but computationally intensive and memory-heavy. RESIRE was developed to provide a high-accuracy, memory-efficient alternative that excels under sparse and noisy data conditions.

The RESIRE Algorithm

RESIRE is formulated as a least-squares optimization problem that minimizes the difference between measured 2D projections and those calculated from the current 3D model. The algorithm iterates through a cycle of forward projection, error calculation, and back-projection to update the model.

Schematic of the RESIRE iterative reconstruction process.

Key technical innovations in RESIRE include:

  • Accurate Projection Steps: The forward step utilizes the Fourier Slice Theorem with oversampling or the Radon transform with sub-voxel precision. The back-projection step is implemented as a mathematically precise linear transformation.
  • Mathematical Convergence: By deriving the Lipschitz constant ($L = nN_z$) for the gradient of the error metric, the authors established an optimal step size for gradient descent, ensuring stable and efficient convergence without manual parameter tuning.
  • Angular Refinement: To account for mechanical imperfections in experimental setups, RESIRE includes a refinement step that optimizes tilt angles within a $\pm 3^\circ$ range.
  • Parallelization: Operating primarily in real space makes RESIRE highly compatible with GPU parallel computing, significantly reducing reconstruction times.

Experimental Results

The performance of RESIRE was validated using both simulated and experimental datasets. In a benchmark test using an experimental aragonite coral sample imaged via ADF-STEM, RESIRE demonstrated significant improvements over traditional methods.

Comparison of tomographic reconstructions of an aragonite coral sample using FBP, SIRT, GENFIRE, and RESIRE.

Quantitatively, RESIRE achieved an R-factor of 5.30% for the coral sample, compared to 13.5% for SIRT and 7.29% for GENFIRE. Qualitatively, the reconstructions show sharper boundaries and a marked reduction in streaking artifacts, particularly in the missing wedge direction (as indicated by the arrows in Figure 4).

Furthermore, RESIRE proved robust in reconstructing extended objects, such as an amorphous tantalum thin film. This capability is crucial for materials science, where samples often extend beyond the imaging field of view. Computationally, the algorithm is highly efficient; a $243^3$ volume can be reconstructed in approximately 30 seconds on a consumer-grade GPU, making it suitable for high-throughput research environments.

Conclusion

RESIRE represents a significant advancement in tomographic reconstruction, offering a powerful combination of accuracy, speed, and mathematical stability. Its ability to handle limited, noisy data and extended objects makes it a versatile tool for a wide range of applications, from atomic-scale materials science to biological imaging. The open-source availability of the MATLAB code ensures that the broader scientific community can benefit from this improved reconstruction framework.

Reference: Pham, M., Yuan, Y., Rana, A. et al. Accurate real space iterative reconstruction (RESIRE) algorithm for tomography. Sci Rep 13, 5624 (2023). https://doi.org/10.1038/s41598-023-31124-7

Summary updated

Abstract

Tomography has made a revolutionary impact on the physical, biological and medical sciences. The mathematical foundation of tomography is to reconstruct a three-dimensional (3D) object from a set of two-dimensional (2D) projections. As the number of projections that can be measured from a sample is usually limited by the tolerable radiation dose and/or the geometric constraint on the tilt range, a main challenge in tomography is to achieve the best possible 3D reconstruction from a limited number of projections with noise. Over the years, a number of tomographic reconstruction methods have been developed including direct inversion, real-space, and Fourier-based iterative algorithms. Here, we report the development of a real-space iterative reconstruction (RESIRE) algorithm for accurate tomographic reconstruction. RESIRE iterates between the update of a reconstructed 3D object and the measured projections using a forward and back projection step. The forward projection step is implemented by the Fourier slice theorem or the Radon transform, and the back projection step by a linear transformation. Our numerical and experimental results demonstrate that RESIRE performs more accurate 3D reconstructions than other existing tomographic algorithms, when there are a limited number of projections with noise. Furthermore, RESIRE can be used to reconstruct the 3D structure of extended objects as demonstrated by the determination of the 3D atomic structure of an amorphous Ta thin film. We expect that RESIRE can be widely employed in the tomography applications in different fields. Finally, to make the method accessible to the general user community, the MATLAB source code of RESIRE and all the simulated and experimental data are available at https://zenodo.org/record/7273314 .
BibTeX
@inproceedings{pham2023accurate,
  author = {Minh Phạm and Yakun Yuan and Arjun Rana and Stanley Osher and Jianwei Miao},
  title = {{Accurate real space iterative reconstruction (RESIRE) algorithm for tomography}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-31124-7},
}
Figure