SRPCA
Disclaimer
This document was primarily created to guide users wishing to reproduce the results from the SRPCA paper: Computational Super-Resolution for Ultrasound Localization Microscopy Through Solving an Inverse Problem (10.1109/TUFFC.2025.3553735). It assumes the user is familiar with the paper material (in particular ULM) and that they are able to run the PALA codes. Furthermore, the use of all the necessary Matlab toolboxes is assumed.
This document reflect the state of the published paper version. If you have any specific questions or concerns, please message me at my gmail address.
If you are using this code, please cite: V. Pustovalov, D. H. Pham, C. Alix, J. -P. Remeniéras and D. Kouamé, "Computational Super-Resolution for Ultrasound Localization Microscopy Through Solving an Inverse Problem," in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 72, no. 5, pp. 636-645, May 2025.
Installation
The SRPCA code was developed to integrate into the pipeline provided by PALA. For the overall code structure and ULM parameters, we refer the reader to the original documentation provided by the authors of PALA (see https://github.com/AChavignon/PALA). However, to make our contribution easier to distinguish from the PALA framework, we have grouped our additions into dedicated folders. Please follow the steps below to integrate the SRPCA code:
- Move the InVivoTours, InVivoPalaSup and AugmentedInSilico folders into the main PALA directory (should be at the same level in the directory tree as PALA_scripts).
- Transfer the contents of the Addons folder into the PALA_addons directory.
- Import the in vivo data into the PALA_data_folder directory.
Data Retrieval
The simulation “In Silico” OPULM PALA dataset is publicly available at: https://zenodo.org/records/4343435. To reproduce the augmented data used in Figure 3 of the paper, one can run the code located in the AugmentedInSilico/GenSilicoAug directory. Alternatively, the pre-generated augmented datasets for various signal-to-noise ratios (SNRs, in dB) are available in the corresponding SilicoData_*dB folders.
The supplementary data to PALA, used for figures 2, 5, and 7 of the paper, can be accessed at: https://zenodo.org/records/7883227. The IQ data can be retrieved from the hdf5 files by using the following Matlab code:
data = h5read(strcat(PALA_data_folder,'IQ_001.hdf5'),'/iq');
IQ = squeeze(data(:,:,:,1))+ 1i * squeeze(data(:,:,:,2));
IQ = permute(IQ,[2,3,1]);The in vivo “WKY” rat brain dataset used in this study is subject to data usage agreements due to institutional constraints. Please contact Jean-Pierre Remeniéras regarding potential data access under appropriate collaborative or institutional agreements.
Reproducing paper results
To facilitate comparison without requiring full re-computation, we provide precomputed result files. Moreover, the MatTrack files, which contain the super-resolved microbubble positions, are available for the SRPCA result on the WKY rat dataset. Please note that MatTrack files were not saved for the in vivo PALA dataset.
Figure 3: simulated “Augmented In Silico” results
The ground truth (GT) in Fig.3.a is directly provided in the PALA data repository and corresponds to the MatOut variable in the PALA_InSilicoFlow_v3_config.mat file.
Use the Fig3b_silicoaug_RSULM.m script to compute Fig.3.b and the Fig3c_silicoaug_SRPCA.m script to compute Fig.3.c. Note that the temporal band-pass filter was not applied to this dataset, as it would have removed a substantial number of microbubbles. The hyperparameters for both RS-ULM and SRPCA were optimized for the 20 dB SNR case and then kept fixed across all other SNR levels for consistency.
Metrics of Table II and Table III can be reproduced by running the SilicoAugMetrics.m code.
Figure 6: In Vivo “WKY” rat results
Although the dataset is not publicly available, we provide the scripts used to generate the corresponding results. The Fig.6.a and Fig.6.c were produced by using the Fig6ac_InVivoTours_RSULM.m script. For Fig. 6b and Fig. 6d, we used the Fig6db_InVivoTours_SRPCA.m script.
Figure 7: PALA supplementary results
To facilitate data handling, we converted the original HDF5 format of this dataset into the standard Matlab .mat format. The complex-valued IQ data were saved using the script described in the section above.
Use the Fig7db_InVivoPala_SRPCA.m script to compute the results of Fig.7.b. and Fig.7.d.
Regarding the results of Fig.7.a. and Fig.7.d. At our request, the first authors of 10.5281/zenodo.7883226 provided unofficial processing scripts and hyperparameters for applying RS-ULM to this dataset, in order to reproduce the figure shown in the associated README file. As noted in https://doi.org/10.1038/s41551-023-01123-0, official reconstruction parameters for ULM are not publicly available. If you wish to access the unofficial code, we recommend contacting the dataset authors directly. While we do not share the scripts themselves, we do provide the results generated using them. However, please note that we cannot guarantee the accuracy or official validity of these outputs.
In Vivo Metrics
Most of the metrics used for in vivo evaluation are straightforward. The Saturation and Gridding score metrics come from the PALA toolbox. The Contrast Ratio and Execution Time are computed using code that we provide.
However, the Fourier Ring Correlation (FRC) metric is more complex to implement, as it is not part of the original evaluation setup. To compute the FRC, we used the notebook and code available at github.com/sakoho81/miplib, which is associated with the publication 10.1038/s41467-019-11024-z. To make the code compatible with ULM images, we modified the main notebook and resolved several compatibility issues to ensure it runs correctly on our system. The adapted version of the code is included in the Addons/FRCmetric directory.
To compute the FRC metric for the in vivo data, first save the MatOut result as a tif image using the PALA WriteTif function and execute the FRCcode.ipynb notebook.