Dynamic laboratory X-ray phase-contrast microtomography with structure-based prior regularisation

TL;DR

This study demonstrates dynamic, lab-based X-ray phase-contrast microtomography by combining free-space propagation XPCI with structure-based prior regularisation to achieve rapid, high-contrast 4D imaging. The method delivers a 9 s temporal resolution at $10.5\,\mu\mathrm{m}$ voxels and a substantial CNR boost (5.8x from XPCI, 29x with regularised phase retrieval) for tracking water uptake in a birch skewer. Phase retrieval converts attenuation images into phase-sensitive data, while the $d\mathrm{TV}$ regularisation anchors time-resolved reconstructions to a high-quality reference, enabling accurate segmentation of dynamic features despite undersampling. The approach opens avenues for plant physiology, pore-scale flow, and in vivo-like dynamic imaging on conventional laboratory X-ray sources, with potential extensions to MI-based reconstructions and higher-resolution setups.

Abstract

X-ray microtomography is a versatile tool allowing the measurement of the 3D structure of optically thick samples. As a non-destructive technique, it is readily adapted to 4D imaging, where a sample can be monitored over time, and especially in conjunction with the application of external stimuli. To apply this technique with the limited X-ray flux available at a conventional laboratory source, we leverage the contrast enhancement of free-space propagation phase-contrast imaging, achieving an increase in contrast-to-noise ratio of 5.8x. Furthermore, we combine this with iterative reconstruction, using regularisation by a structure-based prior from a high-quality reference scan of the object. This combination of phase-contrast imaging and iterative reconstruction leads to a 29.2x improvement in contrast-to-noise ratio compared to the conventional reconstruction. This enables fully dynamic X-ray microtomography, with a temporal resolution of 9 s at a voxel size of 10.5 $μ$m. We use this to measure the movement of a waterfront in the fine vessels of a wooden skewer, as a representative example of dynamic system evolving on the scale of tens of seconds.

Figures (7)

• Figure 1: Numerical example showing a slice through an initial volume (a) that transforms through some dynamic process into a new state (b). The corresponding vectors describing the absolute direction of the image gradients in a and b are illustrated in c and d respectively. Because of the temporal evolution, the slices through the volume in (a) and (b) appear different, however the mutual information encoded in the normalised gradient directions remains constant.
• Figure 2: Comparison of the same birch wood skewer imaged with a propagation distance of $R_2 \approx$ 0 mm (a) and $R_2 =$ 140 mm (b). Plotted line profiles illustrate increased structure and phase fringes at the sample edges. Both images are the average of 5 x 3 s exposures, indicating an increase in sample contrast at the same exposure.
• Figure 3: Axial slices of the reconstructed wooden skewer during the 9 s temporal resolution dynamic scan, conventional (a), phase-retrieved (b), and regularised phase-retrieved (c). Axial slice of the high-quality reference reconstruction, which was used as a structure-based prior for the regularisation (d). Line profiles (e) from the slices in (a) and (c) demonstrate the much greater CNR of the regularised phase-retrieved reconstruction, despite using the same raw data as the conventional reconstruction.
• Figure 4: Segmented axial slice of the regularised phase-retrieved reconstruction (a). Red-bordered regions indicate unfilled vessels, while blue-bordered regions indicate water filled vessels. The green rectangles indicate regions of the wood used to measure the distribution of grey-values. Histograms of voxels falling into the water, air, or wood categories are shown for the conventional (b), the phase-retrieved (c), and the regularised phase-retrieved reconstruction (d).
• Figure 5: A coronal slice of the regularised phase-retrieved volume $u(t = 9 s)$. Zoomed regions are indicated on the figure, illustrating structure within the skewer (b), and the water level within the container (e). Edge response functions are taken (c) to characterise the spatial resolution in a static region of the regularised phase-retrieved reconstruction (b) in red, a moving region of the same reconstruction (e) in blue, and the equivalent moving region in the conventional reconstruction (d) in green. The highest spatial resolution is observed in the static case where it matched the theoretical resolution of the imaging system. The regularisation vastly improves on the SNR of dynamic regions, however with a threefold compromise on the achieved spatial resolution.