FBP3D (ToMoBAR)

Description

Filtered Back-Projection (FBP) method is implemented in ToMoBAR software [5] and exposed in HTTomolibGPU. The method consists of two parts, filtration of the projection data in Fourier space using a custom built SINC filter and then back-projecting the filtered data using ASTRA-Toolbox. The filtering part is implemented using CuPy API and back-projection is a ray tracing operation on the GPU using ASTRA [9]. Notably the filtered data is passed to the back-projection routine directly, i.e., bypassing the device-host transfer, which makes this method more efficient on the GPUs, compared to FBP2D (ASTRA-Toolbox).

Where and how to use it:

Together with Log-Polar 3D (ToMoBAR), this can be a first choice reconstruction method in the library. It is also fast and in well-sampled data situations delivers a similar reconstruction quality as the Log-Polar reconstruction method. It is also recommended to use the FBP method when the data is not well and/or unevenly sampled. Note that for that type of data it might be better to to use iterative algorithms instead, such as CGLS 3D (ToMoBAR) or more advanced FISTA 3D (ToMoBAR).

What are the adjustable parameters:

Most of parameters are self-explanatory from the method’s API httomolibgpu.recon.algorithm.FBP3d_tomobar, so we will mention only the ones that potentially need more explanation.

• detector_pad This parameter is responsible for padding of both (left/right) sides of the horizontal detector of each radiograph. This type of padding extend the edge assuming the sample outside the field of view is approximately similar. This should be used in cases when the sample is larger than the field of view or when the halo-type artifacts should be minimised. Set it to True to enable automatic padding to avoid artifacts (middle image bellow).

• recon_mask_radius This applies the circular mask to the reconstructed volume by zeroing all data outside a certain radius (right image bellow).

Fig. 42 FBP3d reconstruction with detector_pad=0. Note the artifacts in the corners of the image.

Fig. 43 Detector padding enabled detector_pad=True. The artifacts were removed.

Fig. 44 Applying circular masking recon_mask_radius=0.9.

• filter_freq_cutoff can behave differently for this method compared to Log-Polar 3D (ToMoBAR). Normally the lower values might give you noisier and sharper reconstruction, while the higher values can lead to a blur. Note that the change of the filter will lead to the change in the dynamic range of the reconstructed image (see colour bars in the figures bellow).

Fig. 45 FBP reconstruction with filter_freq_cutoff = 0.3 (default value).

Fig. 46 FBP reconstruction with filter_freq_cutoff = 2.0. A slight loss of the resolution, yet reduced noise.

Practical example: