# -*- coding: utf-8 -*-
# Copyright 2019 - 2024 United Kingdom Research and Innovation
# Copyright 2019 - 2024 Technical University of Denmark
# Copyright 2019 - 2022 The University of Manchester
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Authored by: Jakob S. Jørgensen (DTU)
# Laura Murgatroyd (UKRI-STFC)
# Margaret Duff (UKRI-STFC)
To further experiment with tomographic reconstruction and data processing using CIL we provide a few additional data sets, suggestions for experiments and links to CIL resources including publications, documentation and demos.
The following data sets have been downloaded and are available on the STFC Cloud shared drive (read only).
Most of the data sets provided are 3D. For faster execution times in the experiments we suggest to extract a single slice and work only with such a 2D data set. From the loaded 3D AcquisitionData object, data3D, a 2D slice AcquisitionData can be extracted using the get_slice method, for example
data2D = data3D.get_slice(vertical=27)
will extract vertical slice number 27 as a new 2D AcquisitionData object, data2D.
For Cone3D cone-beam datasets only the central slice can be extracted as a Cone2D (fan-beam) dataset, this is done using
data2D = data3D.get_slice(vertical='centre')
If an even number of slices are available in data3D, then the above command will produce a slice that is interpolated from the two central slices.
Nikon datasets can be loaded using the CIL NikonDataReader in the cil.io module.
Zeiss datasets can be loaded using the CIL ZEISSDataReader in the cil.io module.
Crystals in clay:
Cone-beam Nikon data set of crystals in clay from, described at
SophiaBeads:
Cone-beam Nikon data set of glass beads. Data sets with different trade-offs between numbers of projections and exposure time are available, the 256-projection data set is available on the STFC Cloud. Description at
LEGO laminography dataset:
A rotary laminography tomography dataset of a sample of LEGO brick acquired on a Nikon micro-CT instrument, described at
HDTomo datasets:
Six Zeiss cone-beam data sets including the walnut dataset, as well as Kinder Surprise chocolate eggs, a USB stick and more, with descriptions at
Sandstone:
The parallel-beam sandstone synchrotron data set, both a small extracted data set and the full projections are available from
FIPS data: A collection of X-ray CT data sets is available from the Finnish Inverse Problems Society:
Try reconstructing different real data sets:
Choose one or more of the datasets listed above. Load the dataset, determine and carry out any preprocessing required, and compute an FBP or FDK reconstruction. Try also to reconstruct using your favourite iterative/regularized reconstruction method. Apply CIL processors to preprocess data as necessary, possibly after having added noise, artifacts or reduced the datasets yourself.
Relevant notebooks: 01_optimisation_gd_fista.ipynb, Week1/01_intro_walnut_conebeam.ipynb, Week1/02_intro_sandstone_parallel_roi.ipynb
Synthetic data:
Try out loading and generating simulated data from phantoms provided by CIL (cil.utilities.dataexample) or the TomoPhantom CIL plugin (cil.plugins.TomoPhantom). Choose either full data or incomplete data of your choice, add noise, and reconstruct first using FBP and then using regularised reconstruction methods.
Relevant notebooks: 01_optimisation_gd_fista.ipynb, 02_tikhonov_block_framework.ipynb, Week1/03_preprocessing.ipynb
Reduced data reconstruction:
Use CIL processors (e.g. Slicer or Binner or Masker/MaskGenerator) to remove parts of or downsample data sets, for example to obtain a reduced number of projections, a limited angle problem, truncated projections (region of interest data), exterior problem, etc. Compare for example different regularised reconstruction methods at increasingly few projections.
Relevant notebooks: 01_optimisation_gd_fista.ipynb, 03_preprocessing.ipynb
Denoising, deblurring, inpainting:
CIL is developed for tomography but can handle general (at present, linear) inverse problems. Using IdentityOperator, BlurringOperator and MaskOperator provided by CIL it is possible to set up denoising, deblurring and inpainting problems. Choose one or more of these problems and a test image, simulate some data, choose a regularised recontruction problem and compute reconstructions.
Relevant notebooks: Week3/01_Color_Processing.ipynb
Effect of regularisation parameter:
Run Tikhonov and TV-regularized reconstruction with a wide range of different values for the regularisation parameter to see the effect on the reconstruction, ranging from under- to over-regularised.
Relevant notebooks: 01_optimisation_gd_fista.ipynb
Anisotropic regularisation:
Normally we use the same regularisation in all spatial dimensions. Sometimes we may have an image with different behaviour in different dimensions, for example smooth in the y-dimension but edges in the x-direction. Using CIL BlockOperators it is possible to use different regularisers in different dimensions, for example a FiniteDifferenceOperator in y and an IndentityOperator (or no regularisation) in x, or FiniteDifferenceOperators in both x and y but having different regularisation parameters. Implement such anisotropic regularisation in a Tikhonov formulation and demonstrate the effect on a synthetic data reconstruction problem of your choice.
Relevant notebooks: 01_optimisation_gd_fista.ipynb
Verify algorithms against each other:
Compare FISTA and PDHG for solving the same problem, such as TV-regularised or L1-norm regularised least squares. As the same optimisation problem is specified, the different algorithms should produce the same solution, when converged. Try to confirm whether they produce the same solution. You may need to run a large number of iterations. You can also compare with the smoothed TV regulariser, in which case the optimisation problem is smooth and so can be solved using the gradient descent algorithm.
Relevant notebooks: 01_optimisation_gd_fista.ipynb
Compare convergence speed of PDHG using different step sizes:
The sigma and tau step sizes in PDHG can have a dramatic influence on the convergence speed. Experiment with different choices (that must satisfy the constraint specified) and compare the convergence speed. Try on different test problems, including synthetic and real data, and see if there is a trend for the best choice of step sizes across data sets, or it is data set dependent.
Relevant notebooks: 03_PDHG.ipynb
SPDHG subsets and probabilities:
In SPDHG we need to specify the number of subsets to use and the probabilities with which to choose each subset and the regulariser. Experiment with different numbers of subsets and probabilities and compare the effect on reconstruction quality and speed.
Relevant notebooks: 04_SPDHG.ipynb
Other algorithms, operators and functions:
Explore other tools offered by CIL such as the LADMM algorithm, TGV regularisation and weighted least squares and Kullback-Leibler divergence data fidelities, set up test problems and try out new algorithms and optimisation problems and compare the results with problems previously solved.
Relevant notebooks: Week3/01_Color_Processing.ipynb
CIL documentation
https://tomographicimaging.github.io/CIL/
Main CIL GitHub repository
https://github.com/TomographicImaging/CIL
CIL demos and training material repository
https://github.com/TomographicImaging/CIL-Demos
Core Imaging Library -- Part I: a versatile Python framework for tomographic imaging
by Jakob S. Jørgensen, Evelina Ametova, Genoveva Burca, Gemma Fardell, Evangelos Papoutsellis, Edoardo Pasca, Kris Thielemans, Martin Turner, Ryan Warr, William R. B. Lionheart, Philip J. Withers
https://arxiv.org/abs/2102.04560
Core Imaging Library -- Part II: Multichannel reconstruction for dynamic and spectral tomography
by Evangelos Papoutsellis, Evelina Ametova, Claire Delplancke, Gemma Fardell, Jakob S. Jørgensen, Edoardo Pasca, Martin Turner, Ryan Warr, William R. B. Lionheart, and Philip J. Withers
https://arxiv.org/abs/2102.06126
Enhanced hyperspectral tomography for bioimaging by spatiospectral reconstruction
by Ryan Warr, Evelina Ametova, Robert J. Cernik, Gemma Fardell, Stephan Handschuh, Jakob S. Jørgensen, Evangelos Papoutsellis, Edoardo Pasca, and Philip J. Withers
https://arxiv.org/abs/2103.04796
Crystalline phase discriminating neutron tomography using advanced reconstruction methods
by Evelina Ametova, Genoveva Burca, Suren Chilingaryan, Gemma Fardell, Jakob S. Jørgensen, Evangelos Papoutsellis, Edoardo Pasca, Ryan Warr, Martin Turner, William R B Lionheart, and Philip J Withers