Improved Acquisition and Reconstruction for Wavelength-Resolved Neutron Tomography

Abstract

Wavelength-resolved neutron tomography (WRNT) is an emerging technique for characterizing samples relevant to the materials sciences in 3D. WRNT studies can be carried out at beam lines in spallation neutron or reactor-based user facilities. Because of the limited availability of experimental time, potential imperfections in the neutron source, or constraints placed on the acquisition time by the type of sample, the data can be extremely noisy resulting in tomographic reconstructions with significant artifacts when standard reconstruction algorithms are used. Furthermore, making a full tomographic measurement even with a low signal-to-noise ratio can take several days, resulting in a long wait time before the user can receive feedback from the experiment when traditional acquisition protocols are used. In this paper, we propose an interlaced scanning technique and combine it with a model-based image reconstruction algorithm to produce high-quality WRNT reconstructions concurrent with the measurements being made. The interlaced scan is designed to acquire data so that successive measurements are more diverse in contrast to typical sequential scanning protocols. The model-based reconstruction algorithm combines a data-fidelity term with a regularization term to formulate the wavelength-resolved reconstruction as minimizing a high-dimensional cost-function. Using an experimental dataset of a magnetite sample acquired over a span of about two days, we demonstrate that our technique can produce high-quality reconstructions even during the experiment compared to traditional acquisition and reconstruction techniques. In summary, the combination of the proposed acquisition strategy with an advanced reconstruction algorithm provides a novel guideline for designing WRNT systems at user facilities.

Keywords: hyperspectral tomography; interlaced scanning; model-based imaging; streaming tomography.

Figure 1

A block diagram of the wavelength-resolved neutron tomography setup at a spallation neutron source. An incident hyper-spectral neutron beam, depicted here with multiple color arrows each corresponding to a neutron wavelength, provides the transmission signal through a sample placed in front of the detector. A time-of-flight (TOF) imaging detector measures the number of neutrons detected as a function of time, which is equivalent to wavelength. These measurements can be binned into wavelengths to obtain a projection measurement corresponding to each wavelength and rotation angle.

Figure 2

Illustration of two ways to acquire tomographic datasets. (a) Representation of the conventional scanning protocol using 6 equally spaced angles between $0^{\circ }$ and ${180}^{\circ }$. (b) Example of one realization of the proposed interlaced scanning protocol that ensures subsequent measurements are far apart ($K=2,N_{\theta }=6$ in Equation (1)). List of view angles using (c) the conventional approach and (d) one realization of the interlaced scan that was used in the experiments in this paper ($K=3,N_{\theta }=36$). The advantage of the interlaced scan is that at any given point during the course of the experiment, the object is scanned fairly uniformly, thereby enabling reconstructions with fewer artifacts that the end users can utilize to make decisions. The parameters of the interlaced scan can be further adjusted so that subsequent measurements are maximally separated (not shown here).

Figure 3

Illustration of the experimental set-up. Two magnetite crystals are mounted on a holder, separated by a piece of cadmium and held by a tape. Each projection image consisted of measurements using a $512\times 512$ micro-channel plate detector which natively binned the images into energy-bins. Each such projection measurement took approximately $1.5$ h of experiment time and the counts per bin in a region of the sample were approximately in the range of 0 to 35 counts as shown in the plot above (average signal strength in a region indicated by the red box).

Figure 4

Cross section from the reconstruction of a single wavelength bin using filtered back projection (FBP) and model-based image reconstruction (MBIR) with different scanning protocols. In each inset, the top row is the FBP reconstruction, the bottom is the MBIR. The left column is the result from the conventional scan, while the right column is from the interlaced scanning strategy. Notice that the MBIR reconstructions are of significantly higher visual quality than FBP. Furthermore, the interlaced scanning provides a clearer image of the sample in a more uniform manner thereby providing useful feedback very early in the scan compared to existing protocols. All images are displayed in the viewing window $\left[-0.001,0.0035\right]$.

Figure 5

Comparison of normalized root mean squared error (NRMSE) and structural-similarity (SSIM) metrics between the current reconstruction and the final reconstruction in Figure 4 using the MBIR algorithm. Notice that the SSIM and NRMSE converge faster for the interlaced scanning approach compared to the conventional scanning protocol.

Figure 6

XY, XZ and YZ cross sections of the final reconstruction corresponding to the first wavelength bin. Notice that the features of the sample are clearly visible in MBIR even though the input data has only 30 projections. The filter and regularization parameters are chosen for approximately similar visual resolution. Notice that the MBIR approach results in a significantly lower noise reconstruction compared to the FBP approach.

Figure 7

Line profile through the center of the rock. Notice that the FBP produces a very noisy reconstruction as a function of wavelength index.