Tomography Bridge Documentation

On this platform you can find the documentation and updates about the tomography-bridge developments.

They are mainly divided in two sections: data acquisition and data processing.

Tomography techniques at ESRF

Considered techniques (T):

1. Absorption tomography / single-distance phase contrast (typically using Paganin’s method or contrast transfer function): usually introducing some propagation distance, after the sample, to let the interference fringes develop. The radiographs are first filtered using Paganin’s method or the contrast transfer function (CTF), and then Filtered Back-Projection (FBP) is used.
2. Holotomography: the sample is usually scanned 4 times, with different propagation distances (as in Ta), and from those 4 distances, phase and amplitude of the wave at each projection angle are recovered. Then FBP is used.
3. XRF-CT: The sample is scanned using a pencil beam, and one or more (usually two multi-element) fluorescence detectors are placed perpendicular to the incoming beam. The fluorescence detectors are single-pixel or fewpixels detectors (each element is a pixel), and they return hyperspectral data (~2000 channels). A fitting procedure (using typically PyMCA) converts the raw spectra into areal mass density projections. After registration of the projections, tomography reconstruction is used to obtain concentration maps.
4. XRD-CT: The sample is scanned using a pencil beam, and an area detector is placed downstream to the sample. The direct beam is blocked by a beam stop. The collected diffraction rings on the flat panel are azimuthally integrated (using e.g. PyFAI), and the resulting hyperspectral sinograms are reconstructed with FBP.
5. Ptycho-tomography: in both near-field and far-field ptychography a projection can be recovered. Regular FBP can be used on these projections after registration. The retrieval of projection images from ptychographic data is not part of this project, thus the description of ptychographic acquisition schemes is beyond the scope of this document.

Scientific and technical problems

Here we list all the identified current and future problems associated with a correct implementation of tomographic processing codes for the above mentioned techniques.

Scientific problems

Ageing and/or recurring scientific problems (SP):

1. Ring artefacts: affecting (Ta) and (Tb) when not doing helical scans or random detector displacement in step-by-step acquisition, and (Td) in continuous rotation when precise beam intensity measurements are not used to correct sinograms (e.g. when using sampling of the diode current instead of integration).
2. Artefacts related to Local tomography: affecting all the mentioned techniques.
3. Drifts and wobbly rotation axes: mainly affecting long scanning and high resolution techniques like (Tb), (Tc) and (Te).
4. Missing wedge: affecting (Tb), (Tc) and (Te) for organic/biologic samples mounted on membranes, and (Ta) when scanning flat samples (e.g. fossils on rock slabs).
5. Probing beam attenuation and self-attenuation: affecting (Tc) and (Td). In (Td) the self-attenuation occurs at the same energy as the probe, whereas in (Tc) it occurs at the lower energies of the X-ray fluorescence emission.
6. Correctly handling multi-element XRF detectors: this only affects (Tc)
7. Crosstalk in XRD’s rings, due to non-negligible sample sizes (parallax effect): this only affects (Td).
8. Scanning techniques, like (Tc) and sometimes (Td), typically reconstruct from few projections and often (for certain elements or phases) noisy data. In general, the other techniques use a large number of projections (FBP suited), but they can be noisy due to dose or temporal resolution constraints.

Current and near future developments, to address current scientific issues (SD):

1. Automation: certain parameters could be identified (semi-)automatically (rotation axis, etc), and certain reconstruction pipelines could be automated, once the parameters have been identified. This reduces the burden on the software users, especially when dealing with many datasets.
2. Diagnostics: mechanics for the detection of problems in the data. There should be ways of detecting possible problems with the acquisitions (or maybe even with the chosen parameters), warn the user, and even in certain cases to propose solutions.
3. Physical corrections: to solve (SPe) for (Tc) and (Td).
4. Shape corrections: to solve (SPg) for (Td).

Other observations (SO):

1. Ring correction strategies to cope with (SPa) already exist. The existing methods either operate on the projection data (namely: projections or sinograms), or on the reconstructed volumes. The integration and maintenance of these methods in the ESRF software have to be guaranteed.
2. To cope with (SPc), as mentioned in the “scanning techniques” report, some alignment techniques and strategies exist and have been implemented by beamline scientists. While the development of new methods could (or should?) involve collaborations with other groups that have worked on this, some traditional methods should be available in the standard ESRF software distribution. The tomo consistency method used for horizontal alignment uses tomographic reconstruction and forward projection.
3. Local tomography problems from (SPb), can sometimes result from multiresolution acquisitions (TPb). As such, their treatment should be addressed in that context. For non multi-resolution local tomography acquisitions, some methods involving padding of the sinograms are used. Some basic support should be present.
4. The solution to the missing wedge problem (SPd) requires usually additional information injected in the reconstruction algorithm. This can range from regularized reconstructions based on TV or wavelet minimizations, to neural networks. The ability to use these methods should be easy, and their integration facilitated (not necessarily as part of the final package for all of them -> plugin system).
5. The different elements of a multi-element XRF detector (from Tc) can be modelled as different detectors in their own right, thus they should be considered as different measurements for the same angle and beam position. The following aspects contribute to different readings of the different elements: the different self-attenuation experienced by the XRF photons in the sample due to the different travelled paths, different effective area of the elements, and the detection noise. The same considerations hold for multidetector setups. Thus, support for multiple detectors should be considered.
6. Multi-channel, multi-modal and time-resolved reconstructions should be considered as a future development some time after the initial release. Thus, their implementation and integration should not be hindered by early architectural choices. Multi-channel reconstructions affect (Tc) and (Td), while time-resolved reconstructions mainly affect (Ta), (Td), and possibly (Tb). Multi-modal reconstructions could potentially mix any of the techniques.
7. The phase fitting/refinement in (Td) could be done either before of after reconstruction, unless the sample size exhibits the cross-talk problem described in (SPg). In this case, it should be performed voxel-wise after a dedicated reconstruction. The ESRF software should not include any fitting capability.

Technical problems

Ageing and/or recurring technical problems (TP):

1. Large data size: affecting mainly (Ta), but also (Tb), (Te) and partially (Td).
2. Multi resolution: the ability to probe the sample with different beam sizes/detector pixel sizes will become wide-spread with the upcoming EBS upgrade. It applies mainly to (Ta), (Tb) and (Td).
3. Multi-scan handling: large objects that exceed the field of view of the detector can be scanned multiple times. This affects mainly (Ta), but it is somewhat similar to the complete projection formation process of (Te).
4. Acquired image distortion: affecting all the techniques using area detectors: (Ta), (Tb), (Td), (Te).
5. Online reconstructions: this feature could be desirable for all the mentioned techniques. Certain factors like long alignment of the projections in (Tc), or the long iterative reconstruction of the projections in (Tb) and (Te) limit its applicability. It remains quite attractive for certain subsets of (Ta) and (Td).

Current and near future developments, to address current technical issues (TD):

1. Compression and data reduction techniques: Compression at the source (raw images), as well as compression of the final volumes. Note: when reading from disk and distributing the load over a cluster, loading and distributing compressed data could reduce the pressure on the available resources.
2. Multi-resolution acquisitions from (TPb) can either be formed by two different acquisitions with different pixel-size and field-of-view (usually for (Ta) and (Td)), or by a single acquisition where a certain region is more finely scanned (future implementation for (Td)). Being the second approach a priority for ID15A, its support should be considered from the start. The ASTRA toolbox currently offers a projector that handles multi-resolution projections. A similar implementation could be explored. A special case is the super-resolution detector from Fraunhofer on BM18. In this case, the projections are acquired with the highest available resolution, and the processing algorithms should return a multi-resolution reconstruction.
3. Point (TPc) could be handled either as a pre-processing adapted stitching, or as a multiple detector projection. Current ID19 solutions are based on the first of the two, and due to their success they should be considered first.
4. Consolidated distortion correction techniques exist. They are available in certain packages like DCT (developed on ID11). Their implementation should be available in the ESRF software.
5. The architectural definition of the ESRF software should be compatible with online reconstructions. Its support should soon be at least enabled for (Ta). If the azimuthal integration speed allows it, it should also be enabled for (Td).
6. Slow acquisition or dose-limited methods often use interlaced acquisitions (multiple scans gradually filling angular space). The irregular angular sampling should be handled by the reconstruction software.

Other observations (TO):

1. Trainings on python, and tutorials on the new framework will be needed. Members of the tomo-bridge could take part to the organization process, to make sure that we get the needed education from DAU.
2. Backwards compatibility with old datasets should be guaranteed. Conversion tools are viable solutions.
3. Compatibility with the current state of the art tomographic data processing tools should be offered. Internal tests against tools like TomoPy should be run, in order to evaluate the ability of our software to offer the same or better level of functionality, and performance. On the other hand, it would allow new users to try out our software solution.
4. The best strategy for the integration of ptychographic acquisitions (Te) is uncertain. The latest developments suggest that the tomographic reconstruction, and ptychographic image reconstruction should be done in a single process, while traditionally the two used to be disjoint. The same applies to the phase retrieval in holotomography (Tb). The first 3D combined approach would have deep consequences on the pre-processing and reconstruction of the data. Thus, it is not a priority, for the moment. The second two-steps approach should be made possible and well integrated in the ESRF software.
5. Distortion correction is comprised of three phases: measuring the distortion, computing the correction, and applying the correction. The first two should only be performed once per optical system. The third is then applied to every image produced with such optical system. This third phase could be considered similar to sample scaling/shrinkage in (Tb), but it has the added complication that it is not a uniform operation.
6. The multi-resolution approach with the Fraunhofer detector should present: a globally lower reconstruction resolution, and a higher resolution in specific regions. This should all be done with the development of dedicated routines.
7. Important to have integration of the software solutions for pre-processing the data (for spectral fits in XRF, azimuthal integration in XRD, phase retrieval, etc) to solve the data processing of a complete tomography experiment. This should incorporate and eliminate BL specific codes.
8. For multi-resolution (TPb), it will be important to have the ability to perform a low resolution scan of type (Ta) or (Tb) and then select high resolution regions on the low resolution reconstruction. This should also be exposed to the other modalities. The expected implementation would allow to perform this operation simply in the acquisition user interface (UI).