Figure 3
Noise2Inverse training procedures for several imaging techniques. (a-c), Static 3D micro-tomography. Acquisition produces a stack of sinograms (a), each of which is split in the angular domain (b). A 2.5D-CNN is supplied with the current input slice and several context slices (shown in white) reconstructed from one part of the sinogram (blue dot). The target slice is reconstructed from a different part of the sinogram (red dot). (d-f), Dynamic micro-tomography of an evolving process. Scans of multiple time steps are acquired (d). With interlaced acquisition, the angular sampling of each time step is slightly displaced with respect to the previous time step (indicated by the “clock’s hands”) (e). Denser angular sampling is achieved by combining the sinograms of multiple time steps (e). A 2.5D-CNN is trained on time steps where the dynamic process has not yet started (f). The input is a reconstruction of a single time step (blue dot), and the target is reconstructed from the sinograms of the other time steps (red and green dots). (g–i), X-ray diffraction computed tomography (XRD-CT). A pencil beam probes a rotating object that is moved sideways in steps (g). Diffraction of the beam gives rise to rings on the detector (in green). Azimuthal integration along rings at different radii yields several sinograms for each slice of the object. These sinograms are split in the domain of the rotation angle (blue and red dots) (h). From these sinograms, a multi-channel reconstruction is computed, representing the diffractogram of the object (in shades of green) (i). Training is performed with multi-channel inputs and multi-channel targets, with inputs and targets reconstructed from distinct parts of the sinograms (blue and red dots).
