Exploring Image Quality Improvements in High-Speed Dual Threshold Photon-Counting Micro-CT

All images were acquired with a high-speed micro-CT system at Excillum in Stockholm, Sweden (see Fig. 1). Using an Excillum MetalJet E1+ X-ray source, which can provide up to 17 times higher flux than conventional solid-anode sources [18] while still operating with a small X-ray spot, scan times were kept in the few-seconds range for a full CT. The X-ray source was operated at 160 kV and 700 W emission power with a 20 μm e-beam focus, which gives an effective X-ray spot of 30 μm. No filtration was added.

The battery samples, a 2170 consumer cell and a 4680 Li-Ion electric vehicle (EV) cell, were placed on a RT100S air-bearing rotation stage (LAB Motion Systems, Belgium) to enable tomographic imaging, with a maximum rotation speed of 3 Hz. Linear stages (Optics Focus, China) enabled sample manipulation along three different axes. The 2170 cell was positioned at a source-object distance (SOD) of 90 mm and the 4680 cell was placed at SOD = 150 mm.

A photon-counting detector, DECTRIS EIGER2 X CdTe 1M-W, with 75 μm pixel size, a 750 μm thick Cadmium Telluride (CdTe) sensor and an active area of 155.1 mm x 38.4 mm [11] was placed at a source-detector distance (SDD) of 468 mm, giving geometric magnification M = 5.2 for the 2170 cell and M = 3.1. A feature of this detector is the ability to acquire images at 2 different thresholds at the same time without affecting the image quality. The imaging settings for the two different batteries are summarized in Table 1.

In total, 3 scans were performed. Each scan was performed with 2 different thresholds with the goal to– after finding ideal settings for a specific use-case– to be able to acquire a scan with 2 optimal thresholds and combine the resulting volume into a single higher quality volume. The rotation stage was triggered to assure the volumes would align after reconstruction and all scans were acquired without removing the battery from setup. The battery was only moved up to acquire flat-field images for each threshold before the corresponding CT scan.

Several corrections were applied to the raw projection images. Dead pixels were removed, before the images were flat-field corrected using open-beam reference images. Pre-processing included interpolation of defective pixels and detector module gaps. For the 4680 cells, a multi-step ring removal was performed according to [19], implemented in Fiji [20]).

Reconstructions were performed using CERA (Siemens Healthineers, Germany) using the FDK cone-beam variant of filtered back-projection [21]. Finally, the reconstructed volume was virtually sectioned to better visualize and measure the signal of the anode/cathode layers and background. The reconstructed volumes are publicly available at [22].

The detector allows for simultaneous readout of two images acquired at different energy thresholds up to a maximum frame rate of 1125 fps for two 16-bit images (2250 Hz for single threshold 16-bit images). Which threshold combinations to choose for the comparison was decided by analyzing the X-ray source spectrum (Fig. 2) as well as the attenuation in different materials of which the battery is composed and the quantum efficiency of the detector. The X-ray source spectrum was simulated using the Monte Carlo software PENELOPE (top blue lines in Fig. 2) and the attenuation of the different battery components (Li-Ion with NMC811–80% Ni, 10% Mn, 10% Co) was calculated to get an estimate of the X-ray spectrum behind the battery (bottom blue line in Fig. 2). The battery attenuation model was built based on [23] and [24].

The contrast between anode (consisting of a 10 μm copper foil with graphite coating, 50 μm on each side) and cathode (a 10 μm aluminium foil with 50 μm coating of NMC811, on each side) should be maximized. The attenuation of these materials was therefore simulated at relevant X-ray energies as shown in Fig. 3. Ideally, to maximize contrast, the difference in attenuation should be as large as possible. The optimal energy range for high attenuation difference would be between 15 and 75 keV. However, as shown by the attenuation simulation (Fig. 2), almost no photons < 40 keV are expected to penetrate the sample. On the other hand, higher energy photons > 100 keV will not contribute to contrast significantly.

Using the known detector response to different energy photons [11] allows to optimize the detector settings to obtain maximum performance. Based on the source and detector characteristics, six different energy thresholds were chosen: Two low-energy thresholds (15 and 20 keV), two medium-energy thresholds (33 and 36 keV) and two high-energy thresholds (50 and 55 keV). The goal with the low-energy thresholds was to maximize statistics and contrast at the trade-off a larger than single-pixel point-spread-function and increased scattering contribution to the signal. As shown in Fig. 2, the 2170 cell is expected to practically absorb all photons below 30 keV and the 4680 cell will absorb almost all photons below 40 keV. Hence, a large number of photons detected in the lower energy bins will be scattered photons or caused by charge sharing. The medium-energy thresholds were chosen above the Cd and Te fluorescence, to ensure improved PSF and spectral threshold separation, and at energies where the detector has approx. 70–80% efficiency (see [11]) and the contrast between the battery components is expected to be good (as indicated in Fig. 3). Finally, for the high-energy thresholds, around half the average energy of the attenuated spectrum was chosen (see Fig. 2), where a significant part of the emitted spectrum will penetrate the sample. The attenuation differences (shown in Fig. 3) are somewhat smaller, but the detector efficiency is still at approx. 70% at 50 keV and decreasing to ca. 40% at 100 keV.

To determine which of the two energy thresholds should be used, and how to combine them, three different methods were used:

All three methods are described in detail below. Method one is performed on the single threshold images, while methods 2 and 3 utilize weighted averages (from 0 to 100% in steps of 10%) of all combinations of 2 thresholds.

The contrast-to-noise (CNR) between the high- and low- attenuating materials of the jelly roll was calculated according to Eq. (1) by fitting the histogram of grey values with a double gaussian. For both battery types and all the investigated energy thresholds, the CNR normalized to the value at 15 keV is plotted in Fig. 7.

For both batteries, the CNR exhibits a dependency on the selected energy thresholds. The CNR for the 2170 battery has a peak at 36 keV which is 10% higher than the CNR measured at 15 keV. Further increases in the energy threshold led to a decrease of the CNR. For the 4680 battery, the CNR behavior is similar but more significant. The peak is broader and flat, and found for energies between 36 keV and 50 keV with a maximum value of 40%. This suggests that an actual absolute maximum higher than 40% might be observed for an energy threshold in between 36 keV and 50 keV. The possibility of further increasing the CNR for the features of interest, stresses the importance of a proper energy threshold selection when working with photon counting detectors.

The varying CNR behaviors between the two batteries are attributed to their distinct compositions and thicknesses. These two parameters affect not only the intensity, but also the hardness of the transmitted spectrum. It can be hypothesized that the CNR maximum for the 4680 battery is found at a higher energy compared to the 2170 because the thicker battery transmits a harder spectrum and absorbs approximatively all the photons below 50 keV (see Fig. 2). Therefore, all the signal detected below 50 keV is due to charge sharing, not by low-energy transmitted photons. According to the results, the CNR highly benefits by neglecting this contribution from the overall signal, and the high 50 keV threshold may compensate as threshold equivalent beam hardening correction.

The CNR (Eq. (2)) was calculated for all possible combinations of low, medium and high energy detector thresholds optimizing the contrast between jelly roll and the rest of the battery. The CNR results are summarized in Fig. 8, for the 2170 cell (left) and the 4680 cell (right).

The combination of well separated thresholds, e.g. 15 and 36 keV or 33 and 50 keV, always seems to result in a maximum of the CNR curve at some ratio of the two parts with up to 50% improvement in CNR compared to using a single threshold image. Which ratio of the combination is most beneficial varies with the different thresholds. Combining 20 keV threshold data with a threshold at medium or high energy typically benefits from a larger contribution from the higher threshold. If the lower threshold is chosen to be 15 keV, or if the combination is a medium- and a high-energy threshold, a more even contribution is favorable for a high CNR. Combining close thresholds, such as 15 and 20 keV, 33 and 36 keV or 50 and 55 keV, provides no or no significant improvement to the CNR, possibly due to similar photon noise behavior at close energies.

An alternative, but more difficult, way of measuring the CNR could therefore be to calculate the CNR between the different layers in the jelly roll. The thinness of the layers, in combination with the high noise level, could however complicate such a measurement.

The average relative contrast between anode and cathode for thresholds 33 and 55 keV (according to Eq. (3)), calculated for different weights from 0 to 100%, is shown in Fig. 9A for a 2170 battery. Figure 9B-D show examples of the image quality for some different weights as marked in Fig. 9A: 100% lower threshold (B), 50% lower/50% upper (C) and 100% upper threshold (D). All anode/cathode contrast curves are summarized in Fig. 11 for the 2170 cell and Fig. 12 for the 4680 cell.

Figure 11 shows the average relative contrast for all threshold combinations for a single virtual slice of 20 μm through a 2170 cell. In Fig. 12, the same data obtained from the 4680 cell is presented. The leftmost figures show the contrast from a combination of the 15 keV threshold images together with 20, 33, 36, 50 and 55 keV threshold data. Data containing the 20 keV threshold is presented in the middle figures, and the remaining data (combinations between medium and high thresholds) is shown in the rightmost figures.

For the 2170 cell (Fig. 11), the highest anode/cathode contrast was measured at a 70/30 ratio of 33 and 36 keV, closely followed by the case of 100% 36 keV data. The contrast was at a minimum when combining 36 and 55 keV data at a 40/60 ratio. Also using only a 20 keV threshold resulted in low contrast. The contrast curves for the 4680 cell (Fig. 12) show quite different trends, but the overall result is similar: Contrast is maximized when reconstructing with 100% 36 keV data. Lowest contrast for the 4680 cell is achieved with a 70/30 combination of 20 and 36 keV or a 60/40 combination of 20 and 50 keV. This highlights the benefits of threshold optimization when dealing with noise limited data. There may not be a benefit combining two noise-limited datasets.

A key limitation of comparing the relative contrast between anode and cathode layers in high-speed photon noise limited scans is the reliance on averaging to obtain a usable signal. Additionally, the data is normalized, hence, even small errors, artefacts, or changes can cause comparably large differences, as shown in Fig. 11 by e.g. combinations containing the 36 keV dataset. Reconstruction artefacts, defective pixels, or slight alignment offsets may have caused these relatively large changes highlighting the limited robustness of this method for noisy data.

Worth noting is that the CNR maximum (see Fig. 8) does not appear to correlate with the calculated maximum anode/cathode contrast (Figs. 11 and 12). This means that optimizing the threshold selection solely for CNR may give a lower contrast between the anode and cathode and hence not lead to any improvement in layer separation. Lastly, summing two close thresholds (15 keV, 20 keV) before or after reconstruction has always proven to be worse than using either of the thresholds or any weighted average of the two.