Modular Algorithmic Frameworks

Modular Algorithmic Frameworks

As the Advanced Photon Source (APS) enters its upgraded era, instruments are producing larger and more complex datasets than ever before. This data scale, along with increased experimental diversity, calls for reconstruction strategies that are modular, distributed, and hardware-aware. Rather than designing single-use algorithms for each modality or constraint, our work focuses on building flexible computational frameworks that can adapt across imaging scenarios and scale with available computing resources. 

At the core of this approach is the Alternating Direction Method of Multipliers (ADMM), which naturally decomposes a global optimization problem into smaller subproblems. These subproblems can be solved independently (on different nodes or GPUs) and then coordinated to reach a consensus solution. This modularity supports flexible model extensions, handles data heterogeneity, and leverages parallelism in modern computing environments. 

Distributed Optimization as the Unifying Strategy

Across several imaging tasks, from tomography to ptychography, we design systems where computation is distributed, and data movement is minimized. 

  • ADMM enables node-local computation, with only limited information exchanged between nodes per iteration
  • Quantized communication schemes reduce bandwidth needs in large cluster deployments
  • Multi-GPU and multi-threaded implementations support reconstructions from high-resolution scans and dynamic experiments 

Modularity for Hybrid and Evolving Models 

Our frameworks are built to accommodate multiple components: physics models, alignment, deformation correction, or priors, as interchangeable modules within the same reconstruction architecture. 

  • Joint phase retrieval and tomography are handled by parallel solvers operating on coupled subspaces
  • Motion correction, alignment, and illumination retrieval are treated as plug-in submodules under the same optimization backbone
  • Regularization methods, learned denoisers, or statistical priors can be integrated without restructuring the core solver 

This structure allows the same framework to be reused across different beamlines or imaging modalities, reducing development overhead and improving reproducibility. 

Alignment with APS Upgrade

The distributed and modular nature of this work aligns with the data-intensive operation modes emerging at APS post-upgrade. These methods are suited for integration with: 

  • High-throughput acquisition pipelines, where reconstruction must keep pace with large volumes of streaming data
  • Experimental setups with geometric or temporal instability, where alignment or deformation modules can be swapped in as needed
  • Beamlines generating coherent or phase-sensitive data, where probe retrieval or joint inversion modules enhance accuracy

By focusing on architecture rather than algorithm, this work provides a flexible and scalable foundation for reconstruction at modern light sources. Each component can be optimized independently and deployed according to the constraints of the beamline, the sample, or the computing environment.

References

  • Miao R, Aslan S, Koyuncu E, Gürsoy D. Distributed Tomographic Reconstruction with Quantization. arXiv preprint arXiv:2410.06106. 2024.
  • Gürsoy D. Direct Coupling of Tomography and Ptychography. Optics Letters. 2017;42(16):3169-72.
  • Gürsoy D, Hong YP, He K, Hujsak K, Yoo S, Chen S, Li Y, Ge M, Miller LM, Chu YS, De Andrade V. Rapid Alignment of Nanotomography Data Using Joint Iterative Reconstruction and Reprojection. Scientific Reports. 2017;7(1):11818.
  • Aslan S, Liu Z, Nikitin V, Bicer T, Leyffer S, Gürsoy D. Joint Ptycho-Tomography with Deep Generative Priors. Machine Learning: Science and Technology. 2021;2(4):045017.
  • Nikitin V, De Andrade V, Slyamov A, Gould BJ, Zhang Y, Sampathkumar V, Kasthuri N, Gürsoy D, De Carlo F. Distributed Optimization for Nonrigid Nano-Tomography. IEEE Transactions on Computational Imaging. 2021;7:272-87.
  • Nikitin V, Carlsson M, Gürsoy D, Mokso R, Cloetens P. X-ray Nano-Holotomography Reconstruction with Simultaneous Probe Retrieval. Optics Express. 2024;32(23):41905-24.
  • Nikitin V, Aslan S, Yao Y, Biçer T, Leyffer S, Mokso R, Gürsoy D. Photon-Limited Ptychography of 3D Objects via Bayesian Reconstruction. OSA Continuum. 2019;2(10):2948-68.
  • Barutcu S, Aslan S, Katsaggelos AK, Gürsoy D. Limited-angle Computed Tomography with Deep Image and Physics Priors. Scientific Reports. 2021;11(1):17740.
  • Majee S, Aslan S, Gürsoy D, Bouman CA. CodEx: a Modular Framework for Joint Temporal Deblurring and Tomographic Reconstruction. IEEE Transactions on Computational Imaging. 2022;8:666-78.