Evaluation of algorithmic requirements for clinical application of material decomposition using a multi-layer flat panel detector

Abstract.

Purpose

The combination of multi-layer flat panel detector (FPDT) X-ray imaging and physics-based material decomposition algorithms allows for the removal of anatomical structures. However, the reliability of these algorithms may be compromised by unaccounted materials or scattered radiation.

Approach

We investigated the two-material decomposition performance of a multi-layer FPDT in the context of 2D chest radiography without and with a 13:1 anti-scatter grid employed. A matrix-based material decomposition (MBMD) (equivalent to weighted logarithmic subtraction), a matrix-based material decomposition with polynomial beam hardening pre-correction (MBMD-PBC), and a projection domain decomposition were evaluated. The decomposition accuracy of simulated data was evaluated by comparing the bone and soft tissue images to the ground truth using the structural similarity index measure (SSIM). Simulation results were supported by experiments using a commercially available triple-layer FPDT retrofitted to a digital X-ray system.

Results

Independent of the selected decomposition algorithm, uncorrected scatter leads to negative bone estimates, resulting in small SSIM values and bone structures to remain visible in soft tissue images. Even with a 13:1 anti-scatter grid employed, bone images continue to show negative bone estimates, and bone structures appear in soft tissue images. Adipose tissue on the contrary has an almost negligible effect.

Conclusions

In a contact scan, scattered radiation leads to negative bone contrast estimates in the bone images and remaining bone contrast in the soft tissue images. Therefore, accurate scatter estimation and correction algorithms are essential when aiming for material decomposition using image data obtained with a multi-layer FPDT.

1. Introduction

In 2D chest radiography, material decomposition techniques can be employed to remove overlying anatomical structures and therefore improve the sensitivity and specificity in the detection of thoracic abnormalities^1–3 such as coronary calcifications^4–9 and pulmonary nodules.^10–13 Recently introduced muti-layer flat panel detectors^13–18 allow for single exposure spectral imaging using standard acquisition parameters. This makes them suitable for easy implementation in clinical routines.

In 2D chest radiography, spectral images are typically decomposed into bone and water bases.² The presence of additional materials in the examined objects, such as adipose tissue, however, contradicts the assumption of a two-material decomposition. This in turn can result in decomposition errors. Moreover, in comparison to a narrow collimated fan beam, as in dual-energy X-ray absorptiometry,^19,20 or the air gap as in computed tomography (CT) and cone-beam CT (CBCT), 2D projection radiographs may feature a significant amount of scattered radiation. This has been demonstrated to potentially compromise the results of material decomposition algorithms.²¹

The objective of this paper is to evaluate the full range of material decomposition algorithms that could be employed on multi-layer FPDT data, ranging from weighted logarithmic subtraction (WLS) (initially introduced for k-edge fluroscopy in combination with quasi-monoenergetic X-ray spectra^22–24) to projection domain decomposition (PDD) techniques²⁵ known from CT. The evaluation focuses on their performance in two-material decomposition under the impact of adipose tissue and scattered radiation, with the goal of determining which algorithm is superior under these conditions. The evaluation aims to define requirements for the necessary correction algorithms.

2. Theory, Methods, and Materials

The objective of material decomposition in 2D chest radiography is to obtain a bone and a soft tissue image from two X-ray projections of the same object that only differ in terms of the detected X-ray spectra.²⁶ The two projections can be acquired using either a kV switching technique with different tube voltages and an optional change in prefiltration or with the use of a multi-layer FPDT. When kV switching is used, the time difference between the two exposures may result in a projection mismatch due to patient movement. In contrast, multi-layer flat panel detectors (FPDTs) provide single acquisition spectral imaging without motion artifacts. However, compared to kV switching data, they typically lack spectral separation.

In 2D projection imaging, the soft tissue image is not equal to a water path length image but is an image where the contrast due to bone is extinguished to match the neighboring water attenuation. Therefore, we propose to refer to the “soft tissue” image as a non-bone contrast (NBC) image instead, similar to a virtual non-calcium (VNCa) image in CT.²⁷

In the case of quantitative material decomposition, the resulting images reveal path length information. When comparing the water path length image to the NBC image, the difference between the two images becomes clear. As bone is removed from the patient volume, voids are created. These voids appear as negative bone contrast in the projected water path length denoted as $l_w$ shown in Fig. 1(a).

Fig. 1.

Difference between the (a) water path length and the (b) NBC image. Bones appear as voids in the water path length image. In the VNBC image, the bone contrast is extinguished.

By summing up the bone path length $l_b$ and the water path length $l_w$, the total path length represented by $l_t=l_b+l_w$ is obtained. This total path length corresponds to the NBC image shown in Fig. 1(b). Other metrics such as areal densities can be derived from the path length information as well.

To improve the sensitivity and specificity in the detection of coronary calcifications or pulmonary nodules, in 2D radiography, quantitatively accurate material decomposition is typically not required. In this case, a bone image with visible bone contrast and negligible water contrast, and an NBC image where bone contrast is smoothly extinguished are sufficient.

2.1. Material Decomposition using Multi-layer FPDTs

Attenuations $q$ of the layer $D$ of a multi-layer FPDT can be described using the following forward model:

$$q_D(\overline{E})=-\ln \left(\frac{\int S(E)g_D(E)\exp (\sum _{i=1}^N-{\mu }_i(E)l_i)\mathrm{d}E}{\int S(E)g_D(E)\mathrm{d}E}\right),$$ (1)

where $S(E)$ denotes the spectrum emitted by the X-ray tube, $g_D(E)$ energy-dependent detector layer response, ${\mu }_i$ the energy-dependent linear attenuation of the $i$th base material, and $l_i$ the thickness of this base material. The detector response $g_D(E)$ of the layer $D$ is calculated as

$$g_D(E)=\prod _{d=1}^{D-1}(\exp (-{\mu }_d(E)l_d))(1-\exp (-{\mu }_D(E)l_D)),$$ (2)

where ${\mu }_D$ is the energy-dependent linear attenuation of the scintillator and $l_D$ is the scintillator thickness.

For decomposition into two base materials ($mat1$ and $mat2$), a minimum of two layers ($top$ and $bot$) is required. This leaves us with two equations:

$$q_{\text{top}}(\overline{E})=-\ln \left(\frac{\int S(E)g_{\text{top}}(E)\exp (-{\mu }_{mat1}(E)l_{mat1}-{\mu }_{mat2}(E)l_{mat2})\mathrm{d}E}{\int S(E)g_{\text{top}}(E)\mathrm{d}E}\right),$$ (3)

$$q_{\text{bot}}(\overline{E})=-\ln \left(\frac{\int S(E)g_{\text{bot}}(E)\exp (-{\mu }_{mat1}(E)l_{mat1}-{\mu }_{mat2}(E)l_{mat2})\mathrm{d}E}{\int S(E)g_{\text{bot}}(E)\mathrm{d}E}\right).$$ (4)

Thus, material decomposition becomes a problem of solving Eqs. (3) and (4) for $l_{mat1}$ and $l_{mat2}$.

2.1.1. Linear material decomposition approach

Assuming that the effective energy $E_{eff,D}$ of the detected spectrum, which equals its mean value, is approximately equal to that of a monoenergetic X-ray beam with the same energy, we can simplify polychromatic attenuations $q$ to monochromatic ones $p$

$$p_{\text{top}}(E_{eff,top})=-\ln \left(\frac{I_{0_{\text{top}}}(E_{eff,bot})\exp (-{\mu }_{mat1}(E_{eff,top})l_{mat1}-{\mu }_{mat2}(E_{eff,\text{top}})l_{mat2})}{I_{0_{\text{top}}}(E_{eff,\text{top}})}\right),$$ (5)

$$p_{bot}(E_{eff,bot})=-\ln \left(\frac{I_{0_{bot}}(E_{eff,bot})\exp (-{\mu }_{mat1}(E_{eff,bot})l_{mat1}-{\mu }_{mat2}(E_{eff,bot})l_{mat2})}{I_{0_{bot}}(E_{eff,bot})}\right),$$ (6)

and the system of linear equations becomes

$$\overrightarrow{p}=\left[\begin{array}{c}{p}_{\text{top}}\\ p_{bot}\end{array}\right]=\left[\begin{array}{cc}{\mu }_{mat1}(E_{eff,\text{top}})& {\mu }_{mat2}(E_{eff,\text{top}})\\ {\mu }_{mat1}(E_{eff,bot})& {\mu }_{mat2}(E_{eff,bot})\end{array}\right]·\left[\begin{array}{c}{l}_{mat1}\\ l_{mat2}\end{array}\right]=\mathbf{M}·\overrightarrow{l}.$$ (7)

Now, the system of linear equations can be solved for $l_{mat1}$ or $l_{mat2}$. Because WLS does not calculate the path lengths directly, but path lengths are scaled by a constant factor $s$,²⁸ we used a matrix-based material decomposition (MBMD) algorithm instead. In the case of MBMD, Eq. (7) is solved for $\overrightarrow{l}$ by inversion of the matrix $\mathbf{M}$.

To improve the material decomposition results, beam hardening correction can be applied prior to MBMD. This algorithm is referred to as matrix-based material decomposition with polynomial beam hardening pre-correction (MBMD-PBC).

To establish a mapping between the polychromatic attenuations $q$ and their monochromatic counterparts $p$, we employ a 2nd order polynomial function as suggested by Stenner et al.²⁹ The polynomial coefficients were determined by computing line integrals with varying bone and water fractions for every possible intersection length and evaluating the mono- and polychromatic attenuation of each line integral.

Because the actual composition of the examined object is unknown, we assume that the object consists of 20% bone and 80% water everywhere.

2.1.2. Nonlinear material decomposition approach

So far, we have discussed a linear approximation of the polychromatic spectrum of an X-ray tube. In contrast, PDD directly estimates material-specific line integrals from dual-energy projections and thereby facilitates implicit beam hardening correction. This is typically achieved by interpolation in precomputed polychromatic look-up tables. ^20,28

Specifically, normalized top and bottom layer attenuation values are calculated according to Eqs. (3) and (4). The attenuation values, along with their corresponding intersection lengths, are stored in material-specific look-up tables. The intersection length $l_{\text{bone}}$ and $l_{\text{water}}$, associated with the measured or simulated attenuation values, are computed by interpolating the entries stored in the respective rows and columns of the look-up tables.

2.2. Simulation Study

2.2.1. Simulation engine

Our simulation engine comprises four essential components: (1) A voxelized digital phantom that integrates anatomical tissues, including bone marrow and adipose tissue. (2) A poly-energetic forward projector that employs Joseph ray-tracing techniques.³⁰ (3) A Monte Carlo-based X-ray scatter estimator which includes scatter kernel smoothing. (4) A realistic detector model including optional quantum noise simulation.³¹ It is assumed that the selected dose level ensures material-specific images are not limited by noise. Consequently, noise-free simulation data are analyzed in this paper, which aids in limiting the number of influencing factors to adipose tissue and scattered radiation only. Semiempirical X-ray source spectra were calculated according to Tucker et al.³² Material densities and mass attenuation coefficients of the phantom tissues and the CsI:Tl scintillator layers were obtained from NIST tables. ³³

2.2.2. Voxelized torso phantom

Voxelized phantoms with varying amounts of adipose tissue were simulated. These phantoms were derived from a CT scan of a commercially available anthropomorphic torso model (CTU-41, Kyoto Kagaku Co., Kyoto, Japan). Initially, a threshold-based segmentation was applied to the CT dataset, classifying each voxel into one of the four materials: Air $<-300\mathrm{HU}$, adipose tissue $-300$ to $-50\mathrm{HU}$, water $-50$ to 250 HU, and mineral bone $>250\mathrm{HU}$. To obtain a more realistic bone model, cancellous bone was modeled as a composite of mineral bone, accounting for trabeculae and bone marrow. The concentration of mineral bone within these voxels was assumed to be directly proportional to the HU and, consequently, to the attenuation values. The voxels with the highest attenuation were assigned a mineral bone fraction of one, whereas the remaining portion of each cancellous bone voxel was modeled as fat. Adipose variability was introduced by wrapping the simulated phantom with adipose layers around the chest, as shown in Fig. 2. We generated phantoms with 0 mm, 20 mm, 40 mm, 60 mm, and 80 mm of adipose tissue added to the front and the back of the phantom.

Fig. 2.

For the simulation study, we generated a digital phantom with a realistic bone structure and different amounts of added adipose tissue. (a) Segmentation of bones and (b) simulation of bone marrow. (c) Segmentation of adipose tissue and (d) 40 mm of adipose tissue added to the front and the back of the simulated phantom.

2.2.3. Simulated acquisition setups

To ensure comparability between simulations and experimental measurements, a triple-layer detector similar to the Reveal™ 35C triple-layer detector¹⁶ was simulated. The thicknesses of the three cesium iodide (CsI) layers were estimated by comparing attenuation measurements of aluminium (Al) and polymethyl methacrylate (PMMA) slabs, with well-defined thickness, at a fixed tube voltage and prefiltration, to corresponding attenuation values derived from ray-tracing simulations. The thicknesses of the three CsI layers were adjusted so that the difference between simulated and measured attenuation values is minimized. The triple-layer CsI stack-up proposed by Lopez Maurino et al.³⁴ was selected as the initial guess. For a 125 kV spectrum modeled according to Tucker et al.,³² the estimated thicknesses of the three CsI layers were 0.26 mm, 0.25 mm, and 0.43 mm for top, mid, and bottom layers. These thicknesses reproduce the attenuation measurements of the 125 kV spectrum but do not necessarily correspond to the exact thicknesses of the detector layers. A schematic drawing of the simulated triple-layer detector is shown in Fig. 3.

Fig. 3.

Schematic drawing of the simulated triple-layer detector (not to scale).

To gain a principled understanding of the impact of adipose tissue and scattered radiation on material decomposition and limit the number of influencing factors, we restrict ourselves to using data from the top and bottom layers only. Consequently, the middle layer acts solely as a filter.

The simulated phantom was placed directly in front of the detector (contact scan). Two typical clinical scenarios were considered: an anteroposterior (AP) bedside chest examination without an anti-scatter grid and a source to image receptor distance (SID) of 115 cm, and a standing position posteroanterior (PA) examination with a 13:1 anti-scatter grid and a SID of 180 cm. Despite the fact that bedside radiographs are typically acquired at tube voltages around 100 kV, as suggested by Maurino et al.,³⁴ we utilized a higher tube voltage of 125 kV with inherent aluminum filtration in both acquisition setups, resulting in a half-value layer (HVL) of 6.36 mm of Al or 0.29 mm of Cu, respectively. This should yield higher counts and thus lower noise in the bottom detector layers.

2.2.4. Evaluation of material decomposition results

Assuming that quantitative accuracy is not required in the two clinical scenarios, we utilized the structural similarity index measure (SSIM)³⁵ to compare image perception. Specifically, we calculated the SSIM between the ground-truth (representing the actual, simulated, intersection length) and the decomposition results obtained using each of the three algorithms. However, it is important to note that SSIM incorporates a luminance term in addition to the contrast and the structural component. This luminance term is influenced by the mean value of the images, which is not desired when comparing non-quantitative images. Consequently, we set the luminance weighting to zero, assuming proper window and level settings during visual comparison of the images.

2.3. Physical Experiments

Experimental image data were acquired using a commercially available Reveal™ 35C triple-layer detector¹⁶ in combination with a Ysio Max digital X-ray system (Siemens Healthineers) and the same anthropomorphic torso phantoms as used in our simulations (“LUNGMAN” and CTU-41, Kyoto Kagaku Co., Kyoto, Japan).

As for simulations, the X-ray tube voltage was set to 125 kV, resulting in a HVL of 6.36 mm of Al or 0.29 mm of Cu, respectively.

3. Results

3.1. Simulation Study

3.1.1. Intensive care unit (ICU) bedside imaging without anti-scatter grid

Preliminary studies³⁶ have indicated a need for scatter correction for bone removal in an ICU bedside imaging configuration that lacks an anti-scatter grid. The impact of scattered radiation gets even more pronounced when reviewing the corresponding bone images shown in Fig. 4.

Fig. 4.

Bone images obtained using (a) MBMD, (b) MBMD-PBC, and (c) PDD in an ICU bedside imaging setup without an anti-scatter grid. Voxelized torso phantom without additional adipose tissue simulated. Results without scatter correction (top row) and with ideal scatter correction (bottom row).

Without scatter correction (top row), bone images exhibit suboptimal decomposition and negative bone contrast across all three decomposition algorithms. Conversely, following ideal scatter correction using the ground-truth scatter obtained using Monte Carlo simulations (bottom row), the bone images exhibit favorable bone contrast. When employing MBMD and MBMD-PBC, residual soft tissue contrast remains visible in the bone images. In contrast, PDD achieves nearly perfect decomposition. These observations are corroborated by the SSIM.

When assessing the SSIM for both the bone and the NBC images obtained using any of the three algorithms across various scatter correction levels (0%, 80%, 90%, 95%, and 100%), a clear trend emerges: improved scatter correction leads to a higher SSIM, whereas inadequate scatter correction results in a decline of the SSIM especially for the bone image. Figure 5 shows the SSIM as a function of the remaining scatter fraction.

Fig. 5.

SSIMs of (a) bone images and (b) NBC images using each of the three decomposition algorithms in the ICU bedside imaging scenario without an anti-scatter grid.

In an ideal scatter correction scenario, PDD demonstrates superior performance compared with MBMD and MBMD-PBC in terms of SSIM for both the bone and the NBC image. However, when residual scatter is present, we observe a more pronounced decline in the SSIM of PDD relative to MBMD and MBMD-PBC. Notably, with more than 5% of remaining scatter, MBMD-PBC outperforms MBMD and PDD in terms of SSIM. The decrease in SSIM is more pronounced in bone images, but it also manifests as visual artifacts in the NBC images. As depicted in Fig. 6, even a remaining scatter of only 5% can lead to visible decomposition artifacts in both the bone and the NBC image.

Fig. 6.

Bone and NBC images with 5% scatter remaining after scatter-correction in the ICU bedside imaging scenario. (a) MBMD, (b) MBMD-PBC, and (c) PDD. Voxelized torso phantom without additional adipose tissue simulated.

As indicated by the SSIM, the decomposition artifacts caused by remaining scatter are more pronounced in the bone images, resulting in negative path length estimates. The scatter-induced artifacts are visually marked by yellow arrows.

Notably, the impact of residual scatter is less pronounced in the bone image obtained using MBMD-PBC. The SSIM stays almost unchanged compared with the case without scatter. However, the reduction in scatter-induced artifacts comes at the expense of residual soft-tissue contrast.

Upon initial visual inspection, the NBC images demonstrate overall accuracy, yet some residual bone contrast remains visible in the spine region behind the heart, also indicated by yellow arrows.

When assessing the SSIMs for different amounts of added adipose tissue, shown in Fig. 7, the principal trend remains the same but at a lower level, compared with the case without added adipose tissue.

Fig. 7.

SSIMs of bone and NBC images obtained using (a) MBMD, (b) MBMD-PBC, and (c) PDD in the ICU bedside imaging scenario without an anti-scatter grid.

Remaining scatter significantly reduces the SSIM values of the bone images. This effect becomes even more pronounced when simulating phantoms with increased amounts of adipose tissue added to both the front and the back of the phantom. In an ideal scatter correction scenario, the SSIM of PDD remains almost unchanged, whereas it drops to values as low as 0.5 for both MBMD and MBMD-PBC. Although 5% remaining scatter only has a negligible impact on MBMD-PBC, this effect diminishes with higher amounts of added adipose tissue.

When the amount of added adipose tissue exceeds 40 mm and the remaining scatter surpasses 10%, the SSIM values for bone images slightly deviate into the negative range. According to the definition of SSIM by Wang et al.,³⁵ negative SSIM values indicate anti-correlation. Although this finding is intriguing, it is important to consider the context: The actual SSIM value is influenced by the constants employed to stabilize the division. Furthermore, in the considered cases, the bone estimation is completely erroneous (negative); therefore, the actual SSIM values are less relevant. Furthermore, Gourdeau et al.³⁷ demonstrated that SSIM can be compromised when images contain negative values, which is the case here.

In the context of NBC images, the impact of scattered radiation on SSIM values is less pronounced but remains noteworthy. Preliminary studies³⁶ have demonstrated that scatter results in bone structures becoming visible in NBC images, thus contributing to the observed decrease in SSIM. In addition, the presence of adipose tissue, combined with uncorrected scatter, contributes to a modest additional reduction of the SSIM. With 100% remaining scatter, PDD exhibits reduced contrast, resulting in further reduced SSIM values.

3.1.2. PA chest imaging with 13:1 anti-scatter grid

Based on the investigation of offline scatter correction in the context of AP chest imaging without an anti-scatter grid, it is expected that physical scatter correction using an anti-scatter grid as in PA chest imaging, while disregarding dose considerations, would improve the decomposition results. However, as depicted in Fig. 8, despite the anti-scatter grid, the influence of scattered radiation remains visible.

Fig. 8.

Bone and NBC images obtained in a PA chest imaging scenario with a 13:1 anti-scatter grid. (a) MBMD, (b) MBMD-PBC, and (c) PDD. Voxelized torso phantom without additional adipose tissue simulated.

The PA chest imaging decomposition results with a 13:1 anti-scatter grid appear similar to the 5% remaining scatter case in the ICU bedside imaging scenario without an anti-scatter grid, shown in Fig. 6. This indicates the need for additional offline scatter correction besides the anti-scatter grid.

Accordingly, Fig. 9 shows the SSIM as a function of residual scatter after offline scatter correction for different amounts of added adipose tissue within the PA chest imaging scenario with 13:1 anti-scatter grid.

Fig. 9.

SSIMs of bone and NBC images obtained using (a) MBMD, (b) MBMD-PBC, and (c) PDD in the standing PA chest examination scenario with 13:1 anti-scatter grid and additional offline scatter correction.

Despite the anti-scatter grid, bone images exhibit high sensitivity to both scattered radiation and adipose tissue. However, notable distinctions emerge among the three algorithms: The SSIM of MBMD remains high in the ideal scenario, where neither adipose tissue is introduced nor scatter correction is compromised. However, the SSIM declines when either adipose tissue is added to the phantom or scatter correction deviates from an ideal correction. The SSIM of MBMD-PBC is lower under ideal conditions and declines when adipose tissue is introduced. Remarkably, even when scatter correction is suboptimal (up to 20% remaining scatter), the SSIM remains almost unchanged. As theoretical considerations suggest, PDD exhibits the highest SSIM of the three algorithms under ideal conditions. When scatter correction is optimal, the SSIM experiences only minimal impact from the introduction of adipose tissue. However, when scatter correction is suboptimal, the SSIM declines to values comparable to those observed in MBMD.

By contrast, the SSIM values of the NBC images obtained using any of the three algorithms are always in the same range and surpass those of the bone images. However, there is a discernible decrease in SSIM due to suboptimal scatter correction and the combination of suboptimal scatter correction and adipose tissue. The most notable divergence among the three algorithms becomes evident under ideal scatter correction conditions. Although adipose tissue results in reduced SSIM values for MBMD, both MBMD-PBC and PDD exhibit only minimal differences in this case.

3.2. Physical Experiments

3.2.1. ICU bedside imaging without anti-scatter grid

Upon comparison of the simulation with measurement data acquired using the Reveal™ 35C triple-layer detector shown in Fig. 10, we observe a high level of agreement between the two.

Fig. 10.

Comparison of bone and NBC images obtained through simulations without scatter correction with those acquired using a Reveal™ 35C triple-layer detector using (a) MBMD, (b) MBMD-PBC, and (c) PDD in ICU bedside imaging without an anti-scatter grid.

Scattered radiation resulting in negative bone contrast estimates and bone images being displayed as “bones black,” leads to a visual similarity between bone and NBC images.

For bone images, simulated and measured contrasts differ slightly: measured bone path-length estimates were slightly higher than their simulated counterparts, particularly for the images obtained using MBMD and MBMD-PBC. By contrast, measured NBC images closely matched the simulations.

3.2.2. PA chest imaging with 13:1 anti-scatter grid

Similar to Fig. 10 for the ICU bedside imaging case, Fig. 11 shows a comparison of simulated standing chest imaging data with measured CTU-41 chest phantom data acquired using the Reveal™ 35C triple-layer detector and a 13:1 anti-scatter grid. Again, a high level of agreement between the two can be observed.

Fig. 11.

Comparison of bone and NBC images obtained through simulations without scatter correction with those acquired using a Reveal™ 35C triple-layer detector using (a) MBMD, (b) MBMD-PBC, and (c) PDD in the standing PA chest examination scenario with 13:1 anti-scatter grid.

Upon visual inspection, simulations and measurements closely agree for both bone and NBC images, with only slight deviations in noise characteristics.

4. Discussion and Conclusion

In this work, we investigated the performance of three different material decomposition algorithms for 2D X-ray imaging based on a triple-layer flat panel detector. Specifically, we investigated material decomposition in the context of two clinical scenarios: an ICU bedside chest examination without an anti-scatter grid, and a standing PA chest examination with a 13:1 anti-scatter grid. In addition to the detector technology that provides spectrally resolved images, the image processing and decomposition framework plays a key role in achieving accurate material decomposition using a multi-layer flat panel detector.

Our simulation study revealed that all three material decomposition algorithms exhibit high sensitivity to residual scattered radiation. In the case of bone images, the presence of adipose tissue, combined with the residual scatter, also impairs the decomposition results.

In ICU bedside imaging without an anti-scatter grid, the high amount of scattered radiation leads to negative bone contrast estimates, rendering bone images unusable. Remarkably, even when scatter reduction reaches 95% the bone contrast in the lower spine area behind the heart remains negative. Although the SSIM serves as a good global metric for measuring overall image quality, it fails to account for localized variations. Certain image regions, such as the ribs in the 95% scatter correction case, may still be usable for diagnostics, whereas others, such as the lower spine area, may not. Consequently, establishing an SSIM threshold above which images are diagnostically usable remains unfeasible. This becomes even clearer in the context of NBC images, where bone structures tend to reappear in the lower spine area, whereas ribs are accurately removed. Therefore, these images may be perfectly usable for the detection of conditions such as pulmonary nodules. Although the influence of adipose tissue is less notable, in combination with residual scatter, it still contributes to an additional reduction in SSIM. For bone images, which are particularly sensitive to both scattered radiation and adipose tissue, this implies that in addition to scatter correction, an estimation of adipose content may also be necessary.

Using a 13:1 anti-scatter grid as in the standing PA chest imaging scenario results in a more accurate decomposition and consequently a higher SSIM even in the case without any offline scatter correction. By comparing this scenario to the ICU bedside imaging scenario without an anti-scatter, we find that the standing PA chest imaging scenario without scatter correction (100% remaining scatter) is roughly equivalent to the 10% remaining scatter case of the ICU bedside imaging scenario. This implies that the findings align with those of the ICU bedside imaging scenario without an anti-scatter grid spanning from 10% remaining scatter down to 0%, which represents an ideal scatter correction.

Initial experimental results agree reasonably well with the simulation results. Discrepancies may arise due to several factors. These include potential mismatches in material composition between the simulated and actual phantoms, especially in the bone model, as well as the difference between the spectrum used for simulation and the one used for measurements. Among other factors such as noise, these variations can lead to different levels of scattered radiation and consequently affect the decomposition results.

One of the limitations of our simulation study is that we did not account for dose and thus noise.

The work presented here aims to demonstrate the fundamental requirements for various algorithms, such as scatter correction algorithms. Consequently, we intentionally avoided delving into individual algorithms, such as kernel-based scatter estimation³⁸ or deep neural networks³⁹ and their subsequent testing.

We employed a relatively simple model of anatomical variation to investigate the essential factors influencing material decomposition using a multi-layer FPDT. Consequently, we intentionally excluded additional anatomical variations, such as variations in areal bone mineral density (aBMD). A more thorough evaluation of the impact of other anatomical variability is best left for future studies using both simulations and measurement data.

In conclusion, we have demonstrated that material decomposition results using a multi-layer FPDT are highly susceptible to interference from uncorrected scatter, regardless of the decomposition algorithm employed. This effect is not limited to the ICU bedside imaging scenario without an anti-scatter grid but also extends to the standing PA chest examination scenario with a 13:1 anti-scatter grid. This highlights the importance of precise scatter estimation and correction algorithms in 2D chest imaging. Although adipose tissue in combination with residual scatter can also slightly degrade the decomposition results, its effect is less pronounced.

Biography

Biographies of the authors are not available.

Disclosures

The authors declare that there are no financial interests, commercial affiliations, or other potential conflicts of interest that could have influenced the objectivity of this research or the writing of this paper.

Code and Data Availability

The data supporting the findings of this study is considered the intellectual property of Siemens Healthineers AG and is therefore not publicly available. However, simulation results are reproducible using publicly available CT datasets and open-source X-ray simulators, such as MC-GPU.

Disclaimer

The concepts and information presented in this paper are based on research and are not commercially available.