Table 1 Details of Noise2Inverse training on the static and dynamic micro-tomography datasets of a fuel cell and X-ray diffraction tomography dataset of a ceramic. The relative section size reports the ratio of the size of the input section relative to the target section. The size of the training volumes varies widely, whereas the number of parameters in the CNN is stable. Reported training times are indicative, and are specific to the hardware.

From: Deep denoising for multi-dimensional synchrotron X-ray tomography without high-quality reference data

 

Static fuel cell

Dynamic fuel cell

XRD-CT ceramic

Noise2Inverse

Channels (input, target)

11, 1

11, 1

11, 11

Relative section size (input : target)

1 : 1

1 : 5

2 : 1

Size

Sinograms (32-bit float)

\(6.3 \hbox { GB}\)

\(342 \hbox { GB}\)

\(9.2 \hbox { MB}\)

Training volume size (voxels)

\(6.86 \cdot 10^{8}\)

\(2.47 \cdot 10^{10}\)

\(2.46 \cdot 10^{6}\)

CNN Parameters

\(5.48 \cdot 10^{4}\)

\(5.48 \cdot 10^{4}\)

\(5.60 \cdot 10^{4}\)

Duration

Training iterations

220,000

171,600

30,000

Training \(+\) Reconstruction duration

\(\sim 20\) hours

\(\sim 43\) hours

\(\sim 90\) minutes