Digital twins enable full-reference quality assessment of photoacoustic image reconstructions

Authors

Janek Gröhl^1,2,3, Leonid Kunyansky⁴, Jenni Poimala⁵, Thomas R. Else^1,2,6, Francesca Di Cecio^1,2, Sarah E. Bohndiek^1,2, Ben T. Cox⁷, and Andreas Hauptmann^8,9

Affiliations

1. Department of Physics, University of Cambridge, U.K.
2. Cancer Research UK Cambridge Institute, University of Cambridge, U.K.
3. ENI-G, a Joint Initiative of the University Medical Center Göttingen and the Max Planck Institute for Multidisciplinary Sciences, Göttingen, Germany
4. Department of Mathematics, University of Arizona, USA
5. Department of Technical Physics, University of Eastern Finland, Finland
6. Department of Bioengineering, Imperial College London, U.K.
7. Department of Medical Physics and Biomedical Engineering, University College London, U.K.
8. Research Unit of Mathematical Sciences, University of Oulu, Finland
9. Department of Computer Science, University College London, U.K.

Summary

Digital twins for photoacoustic imaging: This repository accompanies the paper, which introduces a digital twin framework to enable full-reference image quality assessment (IQA) in photoacoustic tomography. In conventional experiments, quantitative comparison of reconstruction algorithms is challenging because no-reference IQA measures are often inadequate, and full-reference measures require a known ideal reference image (which is unavailable for real tissue or physical phantoms). The paper’s approach overcomes this by using numerical tissue-mimicking phantoms and a model of the imaging system as a digital twin of the experiment. By calibrating the simulations to match experimental data, the authors create a reference object that is close to a “ground truth” and corresponding simulated sensor data that mimic the real experiment, thus reducing the simulation-experiment gap.

Using this digital twin, the paper quantitatively compares multiple state-of-the-art image reconstruction algorithms for photoacoustic imaging. Among these is a fast Fourier transform-based reconstruction algorithm for circular detection geometries, which is tested on experimental data for the first time. The results demonstrate that the digital phantom twin approach enables rigorous, full-reference IQA of reconstructions: for example, the Fourier algorithm achieved image quality comparable to iterative time reversal but at significantly lower computational cost. This highlights the utility of the digital twin framework for assessing reconstruction accuracy and the fidelity of the forward model, facilitating fair comparisons of different algorithms.

Code Overview

The code contains Python scripts and modules implementing the image reconstruction algorithms and supporting utilities for data processing and analysis. The repository includes multiple reconstruction methods: Delay and Sum, Filtered back-projection, model-based reconstruction (MB), and a fast Fourier transform-based (FFT) reconstruction algorithm for circular sensor geometries. Each method has a dedicated script or function in code/recon_algorithms for performing the reconstruction on the input data.

• Delay and Sum: code/recon_algorithms/backprojection.py
• Filtered back-projection: code/recon_algorithms/backprojection_hilbert.py
• Model-based reconstruction (MB): code/recon_algorithms/modelbased.py
• Fast Fourier transform-based (FFT) reconstruction: code/recon_algorithms/fft/fast_inverse_CRUK.py

Caution

Please note that the time reversal (TR) and iterative time reversal (ITTR) algorithms are not included in this repository, but the reconstruction results are included in the data on Zenodo to allow reproducing the results.

Configuration of paths and parameters is centralised in utils/constants.py. Before running any reconstruction, open this file and adjust the file paths to point to your local data/ directory. This ensures the code knows where to find the input files and where to save outputs.

Typical usage workflow:

1. Run reconstruction scripts: Execute the scripts for each reconstruction algorithm to generate reconstructed images by running the files listed above. For example, running the FFT reconstruction script will read the time-series data from data/.../raw/ and produce reconstructed images in data/.../recons/fft/. Similarly, run the BP, TR, ITTR, MB, and BPH reconstruction scripts to populate their respective subfolders. Each script uses the experimental (exp) or simulated (sim/sim_raw) data as input and saves the reconstructed image files (e.g., as NumPy arrays or images) in the corresponding algorithm’s folder under recons.
2. Compute IQA metrics: After all reconstructions are obtained, use the provided compute_measures.py script to calculate full-reference quality metrics. This step will compare each reconstructed image against the ground truth image from p0/ (the “reference” provided by the digital twin) and compute metrics such as PSNR, SSIM, and other figures of merit. The metrics can be computed for both the simulated data (where ground truth is known) and, by extension, help evaluate how well the calibration holds for experimental reconstructions. The results of this step might be saved as a table or CSV, or printed to the console, depending on the implementation. By changing the data_source parameter from (exp) to (sim/sim_raw), it can be controlled whether the measures should be computed on the experimental or simulated data sets.
3. Visualisation: Finally, run the visualisation or plotting scripts to generate figures that compare the different reconstruction methods. These scripts can create side-by-side image comparisons, difference images, or plots of the IQA metrics for each algorithm. This helps reproduce the figures and quantitative comparisons presented in the paper. For instance, you might generate a figure showing all reconstructions for a given phantom, or a bar chart of the SSIM/PSNR values of all methods on the testing set. There are dedicated scripts for each figure of the manuscript:

• compare_reconstructions.py - Figure 4 in the manuscript
• full_view_vs_limited_view.py - Figure 5 Step A: First run this for model-based, iterative time reversal and FFT-based reconstruction.
• full_view_vs_limited_view_overview.py - Figure 5 Step B: Afterwards, this script reproduces Figure 5 in the manuscript.
• show_forward_model_accuracy.py - Figure 3 in the manuscript

Citation

If you use this code or data in your research, please cite the corresponding paper:

Janek Gröhl et al., “Digital twins enable full-reference quality assessment of photoacoustic image reconstructions,” arXiv preprint arXiv:2505.24514 (2025).

Data

All data are publicly available on Zenodo: https://doi.org/10.5281/zenodo.15388429

License

This project is released under the MIT License (see the LICENSE file for details).

Acknowledgements

Funding Acknowledgements: Development of this code and the underlying research were supported by multiple grants:

• Deutsche Forschungsgemeinschaft (DFG) – projects GR 5824/1 and GR 5824/2 (supporting J.G.).
• U.S. National Science Foundation (NSF) – award DMS-2405348 (supporting L.K.).
• Cancer Research UK (CRUK) – A29580 (supporting T.R.E.).
• Research Council of Finland – Flagship programme projects 359186 and 358944, Centre of Excellence projects 353093 and 353086, and Academy Research Fellow project 338408 (supporting A.H. and J.P.).
• European Research Council (ERC) – European Union Horizon 2020 programme, grant 101001417 (QUANTOM project, supporting J.P.).
• Finnish Ministry of Education and Culture – support for a doctoral programme pilot “Mathematics of Sensing, Imaging and Modelling” (supporting J.P.).
• Engineering and Physical Sciences Research Council (EPSRC), UK – grants EP/W029324/1 and EP/T014369/1 (supporting B.T.C.), as well as EP/R014604/1, which provided support for the Isaac Newton Institute’s “Rich and Nonlinear Tomography” programme.
• Physics of Life Network Phase 3 (PoLNET3) – summer 2023 student bursary “Calibrating numerical photoacoustic forward models with experimental measurements” (supporting F.D.C.).