Research data for dissertation 'Real-time Tomographic Reconstruction'

This repository contains research data and software used for the research presented in the dissertation "Real-time tomographic reconstruction" by Jan-Willem Buurlage. Made permanently available as required by the research data policy of Leiden University.

*Chapter 2*: A modern interface for BSP programs

This chapter outlines the design and use of a new interface for bulk-synchronous parallel programs called Bulk. The 'chapter2/' directory programs contain source code for the three numerical experiments used to verify Bulk: a benchmarking program 'benchmark.cpp' that measures the BSP parameters, source code for the fast-fourier transform 'fft.cpp' and a comparison to a BSPlib implementation 'fft_bspedupack.cpp'. The Bulk library itself is also included.

*Chapter 3*: Geometric partitioning for tomography

This chapter describes a partitioning method that is based on the acquisition geometry underlying the tomographic system matrix, rather than the nonzero pattern. The main results are presented in four large tables, which can be generated using the programs implemented in 'chapter3/src/tableX.cpp'. The main dependency is the "TPT" which was developed alongside this research, and is also included in the directory. The tasks have high computational cost. The runtime measurements were performed on the Lisa supercomputer of SURFsara.

The acquisition geometries used are represented in TOML files in 'data/geometries'.

*Chapter 4*: A projection-based partitioning method

This chapter describes a refinement of the previous method, that is based on a continuous model of the projections that make up the acquisition geometry. The programs used to generate the numerical results as well as most of the illustrations are contained in the directory 'chapter4/'. In particular (files relative to this directory):

- Figure 4.3 and 4.4: generated by 'src/overlap.cpp'
- Figure 4.5: generaed by 'python/plot_partitioning_blender.py'
- Table 4.1 and 4.2: generated by 'src/grcb.cpp' and 'TPT/tools/generate_comvol_table.cpp'
- Figure 4.6: Generated by running 'Pleiades/src/reconstruct.cpp' on a GPU
  cluster of 8 nodes with a 40 Gbit Mellanox Infiniband connection. Each node
  has four NVIDIA GeForce GTX TITAN X GPUs, two Intel Xeon E5-2630 v3 CPUs
  running at 2.40GHz, and 128GB RAM

The acquisition geometries used are represented in TOML files in 'data/geometries', and were used as input to the programs above.

*Chapter 5*: Real-time quasi-3D tomographic reconstruction

This chapter describes a new reconstruction framework and software package based around the idea of reconstructing arbitrarily oriented slices that can be adjusted on the fly.

- All results in the paper were generated directly using the RECAST3D software, and the code snippets are directly given as examples in the corresponding software folder.
- The benchmark results in Table 5.1 were produced on a node with two Intel Xeon E5-2623v3 processors, 128 GB RAM, and two dual-GPU NVIDIA GTX TITAN Z cards for a total of 4 GPUs with 6GB RAM each. To obtain these numbers, 'RECAST3D/slicerecon/src/slicerecon_server.cpp' should be run with a '--bench' flag.
- The data used for the experimental verification, presented in Figure 5.8, can be found at .

*Chapter 6*: Application of quasi-3D reconstruction to synchrotron tomography

This chapter describes an application of the RECAST3D software to synchrotron tomography. The results were generated using the RECAST3D software, as explained above. Adapters to make it work together with the GigaFRoST camera were written in Python, and a slightly modified version of RECAST3D was used for generating some of the results presented in the paper. Later, these changes were merged back into RECAST3D. The original code, which was used to generated some of the results shown in the chapter, is given in the Zipfile 'chapter6/tomcat-live-master.zip'.