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 [10]. 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 is a first choice method in the HTTomolibGPU 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-sampled or unevenly sampled. But in that scenario, however, it is recommended to use iterative algorithms instead.

What are the adjustable parameters:

• On detector_pad and recon_mask_radius parameters see Log-Polar 3D (ToMoBAR) as they produce the same behaviour.

• 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. 43 FBP reconstruction with filter_freq_cutoff = 0.3 (default value).

Fig. 44 FBP reconstruction with filter_freq_cutoff = 2.0. A slight loss of the resolution, yet reduced noise.

Practical example: