Introduction

As energy demands continue to increase into the 21st century, the need for design and deployment of efficient, clean energy sources becomes more urgent. The next generation of nuclear reactors is currently under design to help meet this need. However, new nuclear fuel forms must undergo rigorous testing and safety review before they can be approved for commercial use. This testing involves a plethora of examination techniques, both non-destructive and destructive [1,2,3,4]. Neutron radiography and tomography are two nondestructive examination techniques that can be utilized to examine fuel forms.

Neutrons primarily interact through absorption and scattering, allowing them to penetrate dense, high Z materials more easily than other particles, such as X-rays. Neutron radiography is a nondestructive examination technique that traditionally uses a film or neutron-sensitive scintillator screen that interacts with neutrons transmitted through a sample to create a two-dimensional image of the sample [5,6,7,8]. Neutrons’ penetration abilities make neutron radiography (and tomography) ideal techniques to examine nuclear fuels, as is done at Idaho National Laboratory (INL) [9, 10] and elsewhere [11,12,13,14,15].

When a set of neutron radiographs is acquired, it can be used to perform computed tomography and produce a three-dimensional image of the sample [14, 16,17,18]. The process of converting a set of radiographs to a tomograph is called reconstruction, and various reconstruction algorithms can be used, each with their advantages and disadvantages [19,20,21,22,23]. Reconstruction performed with different algorithms may have different artifacts, noise levels, and emphasize different features.

Image fusion is when data from multiple streams are combined into a single data set, providing additional information to the original separate data sets. It can be performed with data from different techniques, such as X-ray and neutron imaging [24,25,26], where each modality provides a different piece of information, or with data collected from the same technique but with different information (location, time, feature emphasis, etc.). Many variations of this technique have been implemented in the medical [27,28,29,30] and security sectors [31, 32]. However, the application of image fusion to material science is still in its infancy.

When neutron tomography is performed on nuclear fuel, features of interest generally include failure mechanisms and locations, such as cracking, bulging, or fuel or cladding relocation. Identifying features of interest and extracting quantitative data from them is crucial to the success of the nondestructive examination as these results inform modeling and destructive examination, amongst other things. This work applies image fusion to neutron tomography data with the goal of improving data quality compared to standard methods.

Experimental

Workflow

Image fusion involves combining two or more datasets to create a single dataset that provides more information than each individual dataset separately. Oftentimes this involves fusing datasets acquired with from two separate modalities. However, in this work two neutron computed tomography data sets are fused together to improve data quality.

First, the sample was fabricated and irradiated at INL’s Transient Reactor Test (TREAT) Facility [33]. This sample was then imaged using neutron tomography and the data was processed to correct for imaging conditions and to reduce noise. The corrected data was then reconstructed using two separate reconstruction algorithms, each with their own advantages and disadvantages. These two separate data sets were then fused together. Figure 1 shows a schematic diagram of the workflow process used in image fusion of neutron images of nuclear fuel. For unirradiated samples, the second step, sample irradiation, would simply be omitted.

Fig. 1
figure 1

Workflow diagram of the experiment and image fusion process

Full size image

SETH E experiment

The SETH E experiment included a fresh fuel rodlet of 4.9% UO2 and zirconium-alloy cladding. The rodlet was housed in the Separate Effects Test Holder (SETH) capsule for the neutron irradiation test and was subjected to gaussian shaped reactivity-initiated accident type transient resulting in an energy deposition of 1312 J/gUO2 and peak measured temperature of 2113 °C in the fuel rodlet. The SETH capsule both holds the rodlet for the irradiation test and provides the desired boundary conditions. The SETH capsule provides separate effects testing due to its inert gas boundary condition around the rodlet which eliminates typical oxidation and related failure phenomena on the outer surface of the cladding. More information about the SETH campaign can be found in [34].

The primary modes of failure were anticipated to be ballooning and melting. Additional details on irradiation on the SETH capsules as well as the post-irradiation examination results, both destructive and nondestructive, have been published in Schulthess, et al. [35]. A SETH capsule is displayed in Fig. 2.

Fig. 2
figure 2

(Left) Optical image of a SETH capsule and (Right) X-ray radiograph of the same SETH capsule. Figure modified from work presented in [35]

Full size image

Neutron radiography (NRAD) reactor

After irradiation at TREAT, SETH E was moved to INL’s Neutron Radiography (NRAD) reactor for neutron tomography measurements [9]. NRAD is a 250 kWth pool type TRIGA reactor with a radial beamline that produces a thermalized fission spectrum. Neutron imaging was performed at NRAD’s North Radiography Station (NRS) which has a neutron beamline diameter of ~ 60 cm, a length-to-diameter (L/D) ratio of 185, and a thermal neutron flux of 4.5 × 106 n/cm2s. The imaging system has a 16 × 10 cm field of view. The capsule was placed in the beamline on a rotation stage and rotated a full 360°. A total of 1081 images were captured, each with an acquisition time of 60 s and a spatial resolution of 60 μm. Figure 3 shows a SETH capsule placed in front of the neutron imaging system.

Fig. 3
figure 3

Optical image of a SETH capsule placed in the neutron beamline against the neutron tomography system

Full size image

Image processing

First, radiographs were open beam and dark image corrected. This was done by acquiring a set of 5 open and dark images, respectively, and then taking a median image of each. A two-dimensional median filter with a neighborhood of 5 pixels × 5 pixels was applied to each corrected radiograph to reduce noise by eliminating hot spots. A processed radiograph is displayed in Fig. 4.

Fig. 4
figure 4

Corrected neutron radiograph of the SETH E capsule

Full size image

Results and discussion

Volumetric reconstruction

Upon collection and processing of neutron radiographs, the set of 1081 two-dimensional images was used for reconstruction into a three-dimensional volume. The Tomographic Iterative GPU-based Reconstruction (TIGRE) toolbox (Version 2.1) [36] was used in conjunction with MATLAB (Version R2020A) to generate the reconstructed volume. The machine running this software was equipped with 2 NVIDIA Quadro RTX GPUs which were used for GPU-based reconstruction. The dataset was reconstructed twice- once using a modified filtered back projection (FBP) algorithm and again with a simultaneous iterative reconstruction technique (SIRT) algorithm.

The Feldkamp, Davis and Kress (FDK) reconstruction algorithm is a modification of the standard filtered back projection algorithm [37, 38]. A standard filtered back projection algorithm is primarily chosen for the relatively quick and computationally inexpensive reconstruction it produces. The FDK modification provides some noise reduction compared to normal FBP [39]. In this work, a Hann filter was applied to the data to reduce noise for the FDK reconstruction.

A SIRT algorithm is an iterative method which iterates between forward and back projection until it converges on a solution. When compared to an FDK reconstruction algorithm, it is more computationally expensive, but provides reduced noise [40]. Figure 5 shows a side view of the reconstructed volumes using the FDK (left) and SIRT (right) algorithms. These images are re-sliced from the complete 3D dataset, which was reconstructed in layers perpendicular to this view. The FDK shows better edges while the SIRT reconstruction has reduced noise and ring artifacts.

Fig. 5
figure 5

(Left) A side view of a slice from the FDK reconstructed dataset and (Right) a side view slice from the SIRT reconstructed dataset

Full size image

Image fusion

Once the FDK and SIRT reconstructions were created, they were then fused together to create a single dataset that utilizes the advantages of both reconstruction algorithms. First, the grayscale values were stretched to better fill the dynamic range of the 16-bit images. This was done using Eq. 1, where DataNew is the image with grayscale filling the dynamic range, DataOld represents the original data, abs is the absolute value function and min and max are the minimum and maximum values of a dataset, respectively.

$$Data_{new}=\frac{Data_{Old}+abs(min(Data_{Old}))}{abs(min(Data_{Old})+max(Data_{Old}))}$$
(1)

The FDK reconstructed data had higher grayscale levels on average than the SIRT reconstructed data. The relative contrasts of the two datasets had to be matched before image fusion was performed, or else the FDK reconstruction (with higher grayscale values) would have contributed too strongly to the combined image. A grayscale correction factor was determined by comparing the values in regions of interests including the fuel region, the thermocouple regions, and the cladding region. The SIRT data was then multiplied by the correction factor to more closely match the grayscale values present in the FDK data.

Once the grayscale values of the datasets were adjusted, a linear combination of the two reconstructions was used to produce the fused data, as described by Eq. 2 where F is the fused image, FDK is the dataset reconstructed by FDK, SIRT is the dataset reconstructed by SIRT, and x is the weighting factor. The weighting factor controls the contribution of the FDK and SIRT reconstructions to the fused image and can be adjusted accordingly to generate a fused image with the desired features from each individual image.

$$F=FDK^*(x)+SIRT^*(1-x)$$
(2)

The results of the image fusion are shown in Fig. 6 where part A of the figure shows the FDK reconstructed slice, part B shows a SIRT reconstructed slice, and part C shows the fused image. A visual inspection reveals that the FDK reconstruction demonstrates pronounced ring artifacts, whereas the SIRT reconstruction tends to blur feature edges. The fused image minimizes both artifacts, providing an improved picture of the SETH E capsule and irradiated fuel rodlet. For this application, a standard deviation filter was used in the weighting factor calculation to emphasize the sharper edge features. Therefore the weighting factor is not a scalar value but varies for each pixel position based on the weighting factor implementation which may be modified from application to application.

Fig. 6
figure 6

A A slice of the reconstructed FDK volume which has significant ring artifacts. B A slice of the reconstructed SIRT volume showing the blurred edges of certain features such as the fuel rod. C A slice of the fused data showing reduced ring artifacts and a sharper fuel edge

Full size image

Analysis

While a visual examination of Fig. 6 can reveal some benefits of image fusion, a more quantitative approach is necessary for a full appreciation of the advantages image fusion can offer. Figure 7 shows a reconstructed slice of the SETH capsule with a yellow line across the fuel region. This shows the location of the line profile plot in the right side of the figure. In the line profile, the grayscale values of the FDK, SIRT, and fused reconstructions are all shown together. As expected, the fused data has grayscale intensities between those of the FDK and SIRT reconstructions because it is a linear combination of the two. The slope of the line profile at the edges of the fuel region (pixel position 20–40 and 200–220) is steeper in the FDK and fused slices, demonstrating increased contrast and less blur at the edges of the fueled region. This is an improvement over the SIRT image. The change in intensity of the grayscale values across the crack in the fuel (pixel position ~ 150) is more pronounced for the FDK and fused images than the SIRT slice, showing that the fused image emphasizes the crack. Lastly the noise seen from ring artifacts (pixel position 100–140) in the fused data is less than that of the FDK slice. These features all demonstrate that the fused image produces higher contrast around features of interest while also minimizing noise such as ring artifacts. These improvements may enable better feature identification and quantification, for example cladding ballooning, and fuel fragmentation and relocation into a ballooned region in the case of the SETH E. Figure 8 shows a rendering of the fused dataset with the broken fuel rod, blistered fuel, and relocated cladding clearly visible.

Fig. 7
figure 7

(Left) A slice of the fused SETH E image with the yellow line showing the location of the line profile. (Right) A line profile of the grayscale values across the fuel rod for the FDK, SIRT, and fused slice

Full size image
Fig. 8
figure 8

 A three-dimensional rendering of the neutron tomography of the SETH E capsule. The broken fuel rod, blistering on the fuel rod, and cladding relocation can all be seen

Full size image

Conclusions

This work demonstrates the improvement in image quality offered by the fusion of data reconstructed with different algorithms. This is particularly useful in neutron imaging applications where quantitative data is extracted as more accurate data can be generated and imaging artifacts reduced. Additional reconstruction algorithms, such as conjugate gradient least squares (CGLS) or simultaneous algebraic reconstructive technique (SART), may be incorporated into future image fusion applications to further reduce reconstruction artifacts or improve image contrast.

Fusion of reconstructions generated from the same measurement was the focus of this work, but fusion of data collected with different techniques can provide unique insights that provide more information than each individual dataset. By correlating thermal and epithermal neutron tomography, X-ray tomography, or gamma emission tomography, a more complete picture of a sample can be achieved. Future work at INL involves leveraging existing tomography capabilities in conjunction with the image fusion methodology presented in this work to generate fused datasets from separate modalities.