Characterization and Potential Applications of a Dual-Layer Flat-Panel Detector

Abstract

Purpose:

Dual energy (DE) x-ray imaging has many clinical applications in radiography, fluoroscopy, and CT. This work characterizes a prototype dual layer (DL) flat panel detector (FPD) and investigates its DE imaging capabilities for applications in 2D radiography/fluoroscopy and quantitative 3D cone-beam CT. Unlike other DE methods like kV switching, a DL FPD obtains DE images from a single exposure, making it robust against patient and system motion.

Methods:

The DL FPD consists of a top layer with a 200 μm-thick CsI scintillator coupled to an amorphous silicon (aSi) FPD of 150 μm pixel size and a bottom layer with a 550 μm thick CsI scintillator coupled to an identical aSi FPD. The two layers are separated by a 1 mm Cu filter to increase spectral separation. Images (43×43 cm² active area) can be read out in 2×2 binning mode (300 μm pixels) at up to 15 frames per second. Detector performance was first characterized by measuring the MTF, NPS, and DQE for the top and bottom layers. For 2D applications, a qualitative study was conducted using an anthropomorphic thorax phantom containing a porcine heart with barium-filled coronary arteries (similar to iodine). Additionally, fluoroscopic lung tumor tracking was investigated by superimposing a moving tumor phantom on the thorax phantom. Tracking accuracies of single energy (SE) and DE fluoroscopy were compared against the ground truth motion of the tumor. For 3D quantitative imaging, a phantom containing water, iodine, and calcium inserts was used to evaluate overall DE material decomposition capabilities. Virtual monoenergetic (VM) images ranging from 40 to 100 keV were generated, and the optimal VM image energy which achieved the highest image uniformity and maximum contrast-to-noise ratio (CNR) was determined.

Results:

The spatial resolution of the top layer was substantially higher than that of the bottom layer (top layer 50% MTF = 2.2 mm⁻¹, bottom layer = 1.2 mm⁻¹). A substantial increase in NNPS and reduction in DQE was observed for the bottom layer mainly due to photon loss within the top layer and Cu filter. For 2D radiographic and fluoroscopic applications, the DL FPD was capable of generating high-quality material-specific images separating soft tissue from bone and barium. For lung tumor tracking, DE fluoroscopy yielded more accurate results than SE fluoroscopy, with an average reduction in the root-mean-square error (RMSE) of over 10×. For the DE CBCT studies, accurate basis material decompositions were obtained. The estimated material densities were 294.68 ± 17.41 and 92.14 ±15.61 mg/ml for the 300 and 100 mg/ml calcium inserts respectively, and 8.93 ± 1.45, 4.72 ± 1.44, and 2.11 ± 1.32 mg/ml for the 10, 5, and 2 mg/ml iodine inserts respectively, with an average error of less than 5%. The optimal VM image energy was found to be 60 keV.

Conclusions:

We characterized a prototype DL FPD and demonstrated its ability to perform accurate single-exposure DE radiography/fluoroscopy and DE-CBCT. The merits of the dual layer detector approach include superior spatial and temporal registration between its constituent images, and less complicated acquisition sequences.

1. Introduction

In recent years, the clinical use of flat panel detector (FPD)-based x-ray imaging has grown rapidly with applications ranging from image-guided radiation therapy^[1] and interventions^[2] to diagnostic imaging utilizing specialized radiographic systems^[3] and dedicated CBCT scanners^[4],[5]. Such FPD-based imaging systems increasingly demand advanced imaging techniques that generate material-specific and quantitative information. Dual-energy (DE) imaging can provide such information and has demonstrated substantial clinical value in 2D radiography/fluoroscopy^[6],[7], and 3D diagnostic CT^[8]. This technique provides improved contrast compared to single-energy imaging by exploiting the dependence of x-ray absorption coefficients on atomic number. DE imaging is especially effective in differentiating soft tissue from other endogenous or exogenous materials that have atomic numbers such as bone or vessels/arteries with iodinated contrast.

DE imaging with a FPD has shown promising clinical applications in many 2D imaging tasks. For example, DE mammography demonstrated improved detection of micro-calcifications by removing the soft-tissue background^[9],[10]. DE chest radiography showed better detection and diagnosis of thoracic abnormalities than single energy radiography^[11] and demonstrated the possibility of offering markerless motion tracking of lung tumors ^[12],[13]; DE x-ray absorptiometry (DXA) allows estimation of bone density by removing the background soft tissue^[14]; DE angiography can enable maskless subtraction angiography^[15],[16]. These promising 2D applications motivated the development of FPD-based DE CBCT systems. A number of preliminary studies demonstrated the benefits of various DE CBCT systems: in diagnostic imaging, Zbijewski et al^[17] implemented DE CBCT for contrast-enhanced musculoskeletal imaging and found >90% accuracy in material classification; in image-guided radiation therapy, Men et al^[18] and Li et al^[19] found DE CBCT improved the CBCT image quality and therefore the accuracy of CBCT-based dose calculation; in interventional radiology, Müller et al^[20] demonstrated the feasibility of DE CBCT on an angiographic C-arm CBCT system.

In many imaging tasks, patient motion is inevitable, including cardiac motion, respiratory motion, and other involuntary movement. In addition, the imaging system may be moving as well, such as in image guidance during a volumetric modulated arc therapy (VMAT) treatment or during a CBCT acquisition. Thus, it is desirable to obtain DE projections in a single exposure to avoid spatial and temporal misalignment, particularly for real-time imaging tasks like lung tumor tracking fluoroscopy^[6]. and digital subtraction angiography^[15],[16]. For CBCT, the system is even more susceptible to motion artifacts due to its longer scan times, so acquiring DE projections in a single rotation would greatly minimize misregistration artifacts.

In diagnostic CT, there are several methods for obtaining DE data: fast kV switching that alternates between low and high tube voltages in consecutive projections^[8], “slow” kV switching that acquires sequential scans at different energies^[21]; dual source systems offset by 90° that simultaneously acquire different projection data at different energies^[8]; and dual layer detectors that simultaneously capture low and high energy projections in a single scan at fixed tube voltage^[8]. The first three methods use different forms of dual-exposure acquisitions, which achieve good energy separation but inevitably introduce temporal and spatial mismatches. On the other hand, a DL detector enables single-exposure DE acquisitions, where the thinner top layer absorbs low energy x-rays and the thicker bottom layer stops the transmitted high energy x-rays. Although it has relatively limited energy separation compared to dual-exposure systems, the DE images are acquired simultaneously and have been shown to be still capable of accurate iodine quantification^[22]. An additional advantage is DE imaging is always available, without needing a special scan protocol. While photon counting detectors can also capture spectral information from a single exposure, they are very expensive, have more limited count rates, and are not yet available in large-area format. To date, most DE imaging reports involving FPDs were proof-of-concept studies and performed using dual-exposure acquisitions. Notable work using layered detectors includes Maurino et al^[23], who proposed a three-layer FPD and demonstrated initial 2D simulation results, which were promising for DE radiography. Another study by Myronakis et al^[24] found that spectral information from a multi-layer MV detector could be obtained, enabling better detection of gold fiducial markers from projection images. These studies, though promising, only demonstrated 2D performance of their layered detectors for spectral imaging tasks.

In this paper, we present the first work investigating the DE imaging performance of a prototype dual-layer FPD (Varex Imaging Corporation, Salt Lake City, UT) for both 2D and 3D kV imaging applications. Some early results were presented at previous conferences^[25][26]. We first characterize the detector by measuring the modulation transfer function (MTF), noise power spectrum (NPS), and detective quantum efficiency (DQE) for both layers. We then explore potential clinical applications by investigating its ability to provide material-specific radiography, real-time motion tracking in fluoroscopy, and quantitative material decomposition in CBCT.

2. Methods and Materials

2.A. Prototype Dual-Layer Flat-Panel Detector

Figure 1 shows a schematic diagram of the prototype dual layer FPD, which was built on a modified Varex XRD 4343RF architecture. The detector comprises two amorphous-silicon (a-Si) panels, each deposited with a CsI scintillator. The scintillator thicknesses for the top and bottom layers are 200 μm and 550 μm, respectively. When a polyenergetic x-ray spectrum enters the detector, the top layer preferentially absorbs low-energy (LE) x-rays while high-energy (HE) x-rays are more likely to penetrate and be absorbed in the bottom layer, generating LE and HE projections in a single exposure. This DL FPD prototype also includes a 1 mm copper filter between the two layers that further improves spectral separation. The layer materials and thicknesses were selected based on available processes and materials (e.g., 200 um CsI deposition for mammography panels, 550 um CsI for standard panels), rather than any specific optimization. The goal was to produce a prototype capable of demonstrating potential applications. However, general principles were applied, including a thinner top layer, a filter to increase spectral separation^[27], and a thicker bottom layer. Optimization of the layers can play an important role in imaging performance (e.g., the filter improves spectral separation at the expense of photon loss in the bottom layer) and is being pursued in ongoing work.

Fig.1.

Schematic diagram of the prototype DL FPD (not to scale).

Each layer has an active area of 432×432 mm², with a native pixel size of 0.150×0.150 mm² (1×1 binning). The dual layer images are combined into a single frame and read out at frame rates up to 7.5 fps for 1×1 binning and 15 fps for 2×2 binning. Both layers have 16-bit depth and provide read-out modes of 7 different gains. We found gain 5 to give suitable sensitivity without saturation in both layers, so it was used for all subsequent experiments. The characteristics of the DL FPD are summarized in Table 1.

Table 1.

Summary of the DL FPD specifications

Parameter	Specifications
Detector type	dual-layer integrated a-Si detector
Top layer scintillator	200 μm CsI:Tl
Top layer pixel size	150 μm
Top layer matrix size	2880 × 2880
Separation between panels	2.5 mm
Bottom layer scintillator	550 μm CsI:Tl
Bottom layer pixel size	150 μm
Bottom layer matrix size	2880 × 2880
Detector size	500 mm × 500 mm × 50 mm
Read-out rate	7.5 fps @ 150 μm (1 × 1) 15 fps @ 300 μm (2 × 2)
Gain selections	7 gain selections for each panel (up to 30:1 range)
Data interface	Optical fiber-channel

2.B. Detector Performance

We first characterized the detector performance by measuring the pre-sampling MTF, NPS, and DQE for both top and bottom layers. Experiments were performed for two x-ray spectra: 1) RQA5, i.e. 70 kV, 50mAs (250 mA and 200 ms pulse width), with 21 mm aluminum (Al) filtration, and 2) 120 kV, 1mAs (100 mA and 10 ms pulse width), with 2 mm external Al filtration. These correspond to a standard beam quality for detector characterization and a typical spectrum for many clinical applications, respectively. Figure 2 shows the entrance spectra at the top and bottom layers for the RQA5 and 120 kV source spectra. Spektr^[28] was used to generate the spectra and analytically attenuate them through the top layer and filter. The resulting spectral separation (difference in average energy between the layers) was 7.5 keV and 24.3 keV, respectively. Without the 1 mm Cu filter, the spectral separation would only be 3.2 keV and 9.5 keV, respectively. All measurements were obtained with 1 × 1 mode at a source-to-imager distance (SID) of 1650 mm. Images obtained from both detector layers were first preprocessed by their respective dark current, gain, and bad pixel corrections. The corrected images were then used for the following characterization tasks.

Fig. 2.

Entrance spectra of RQA5 and 120 kV for the top and bottom layers.

2.B.1. Pre-sampling MTF

Pre-sampling MTF was obtained using a 1 mm thick tungsten edge phantom, which was placed on the center of the detector surface at a small angle (3°) with respect to the vertical axis. The slightly angled placement allowed the acquisition of an oversampled 1D edge spread function (ESF) along the length of the edge. For both detector layers, an oversampling factor of 2 was used to resample the ESF to a fixed bin size of 0.075 mm. The rebinned ESF was differentiated to yield a line spread function (LSF), and the pre-sampling MTF was estimated as the result of the Fourier transform of the LSF. This MTF was normalized by its zero-frequency value to obtain the final MTF.

2.B.2. Normalized NPS

A set of flat-field images with 100 frames was acquired for NPS estimation. An area of 20×20 cm² at the center of the flat-field image was selected and combined for all the frames for NPS analysis. Each area was further broken up into a series of half-overlapping 256×256 sub regions of interest (ROIs), denoted as I_ROI. The normalized NPS (NNPS) is calculated as:

$$\mathit{\text{NNPS}}\left(u,v\right)= \frac{\Delta x\Delta y}{M\cdot 256\cdot 256}\sum _{ROI=1}^M{|DF{T}_{2D}\left\{\frac{I_{ROI}\left(x,y\right)-\overline{I_{ROI}}}{\overline{I_{ROI}}}\right\}|}^2,$$ (1)

where u and v are the spatial frequencies along the x and y directions, respectively, Δx and Δy are the pixel pitch, $\overline{I_{ROI}}$ is the mean signal intensity of the ROI, and M is the total number of I_ROI. The final 1D NNPS is the average result of 1D NNPS along the central u and v directions.

2.B.3. Detective quantum efficiency

The DQE measures the efficiency of a detector in converting incident x-rays into images. It is generally defined as:

$$DQE\left(u,v\right)= \frac{SN{R}_{out}^2}{SN{R}_{in}^2}= \frac{MT{F}^2\left(u,v\right)}{\mathit{\text{NNPS}}\cdot \left(\frac{\varphi }{ K_a}\times \mathit{\text{Dose}}\right)},$$ (2)

where $\frac{\varphi }{{ K}_a}$ is the fluence per air kerma for the RQA5 and 120 kV spectra. The Dose was measured using a Raysafe X2 solid-state x-ray detector (Unfors Raysafe AB, Billdal, Sweden) at a distance of 1300 mm from the x-ray source and then scaled by the inverse-square law to obtain the dose at the detector surface. The Raysafe dosimeter had a valid calibration from a dedicated dosimeter calibration lab at the time of this work.

2.C. 2D imaging applications

2D DE tasks involve acquisitions of two projections at different energies, which can be used to generate material-specific projections that better characterize pathology or provide other information. For example, DE chest radiography was previously shown to improve the accuracy of detecting and diagnosing thoracic abnormalities. In this work, we used an anthropomorphic thorax phantom (The Phantom Laboratory, Salem, NY, USA) containing a porcine heart with coronary arteries filled with a barium-impregnated gel to evaluate the DL FPD’s ability to perform material-specific radiography. The barium concentration was roughly equivalent to undiluted Iohexol contrast agent (350 mg/cm³ iodine)^[29]. The DE projections were acquired with a single exposure of 120 kV, 8 mA, and 18 ms on a tabletop system, with a SID and source-to-axis distance (SAD) of 1300 mm and 800 mm, respectively. The experimental setup is shown in Fig. 3(a), where the DL detector and the x-ray source were mounted on the tabletop, and the thorax phantom was placed on a rotating stage, which enabled CBCT acquisition. The rotating function was disabled for 2D imaging tasks.

Fig. 3.

(a) Experimental setup for 2D imaging acquisitions. (b) Motion phantom setup.

Due to the stacked design of the DL FPD, a one-time geometric calibration was needed to compensate for the slight differences in depth and position between the two layers. The geometric calibration was performed using an IsoCal phantom (Varex Imaging Corporation, Salt Lake City, UT), which is commonly used to measure the relative locations of the source, detector, and rotation axis needed for 3D reconstruction. We chose this particular phantom because its projection image included highly attenuating BBs distributed across the entire detector FOV, which was ideal for whole-field calibration. The HE projection (bottom layer) was registered to the LE projection (top layer) using an affine transformation to account for scaling, rotation, and translation. The resulting affine transformation was then used to register the layers in all future scans.

2.C.1. Scatter correction

Scatter adds undesired signal and noise to the images, potentially causing errors when performing material decomposition tasks. Thus, we first performed a scatter correction for both layers using CBCT Software Tools (CST 2.0, Varex Imaging). The software estimates the scatter distribution using a fast adaptive scatter kernel superposition (fASKS) method^[30]. Among numerous scatter correction methods, the fASKS method has the advantages of being computationally efficient and free of extra hardware such as beam blockers. Due to differences in our geometry, spectrum, and detector, we applied a global scaling factor to the fASKS estimates. The scaling factor was determined by a one-time measurement of the scatter in the shadow of a 3 cm diameter, 2 mm thick lead disc placed on the entrance side of the phantom. The estimated scatter distributions were then subtracted from the raw projections for scatter correction.

2.C.2. Dual energy radiography

After log normalization of the scatter-corrected LE and HE projections, material-specific projections were generated by weighted subtraction of the LE and HE line integrals, denoted as l_L and l_H. The soft tissue and bone projections, l_s and l_b, are expressed as:

$$\begin{gathered} l_s=l_H- \frac{{\mu }_b^H}{{\mu }_b^L} l_L , \\ {l}_b=-l_H+ \frac{{\mu }_s^H}{{\mu }_s^L} l_L, \end{gathered}$$ (3)

where ${\mu }_b^H,\mathrm{ }{\mu }_b^L$ and ${\mu }_s^H,\mathrm{ }{\mu }_s^L$ are the effective attenuation coefficients of bone and soft tissue for the HE and LE spectra, respectively.

2.C.3. Tumor tracking using DE fluoroscopy

2.C.3.a. Experiment setup

The phantom and setup used in 2.C.1 were used for this task. We chose to use a grape to simulate a tumor as it has similar composition and shape to a lung nodule. The grape was cut into a semi-ellipsoid shape with a cross section diameter of 20 mm and thickness of 10 mm. As shown in Fig. 3(b), the tumor was placed between the thorax phantom and the detector surface, resulting in a magnification factor of 1.34. To simulate lung tumor motion during respiration, the simulated tumor was connected to, and driven by, a Faulhaber LM1247-060-11 Quickshaft linear DC-servomotor and MCLM3006 motion controller (Faulhaber GmbH, Schönaich, Germany) producing respiratory motion in the inferior-superior direction. The respiratory motion trajectory was defined by an analog sine waveform connected to the analog input of the motion controller. The signal was created using a Rigol DG1022 arbitrary waveform generator (Rigol Technologies, Suzhou, China). Note that the sine wave might not correspond to realistic respiratory motion, though our goal is to demonstrate the tumor tracking ability of this prototype detector. Realistic respiratory models will be included in our ongoing work.

The DE fluoroscopy projections were acquired at 120 kV, 15 mA, and 18 ms. The simulated respiratory motion followed the sine wave programmed with a peak-to-peak amplitude of 20 mm and a period of 5 sec. The fluoroscopy projections were obtained with detector readout rate of 15 fps for 2 period cycles (i.e., 10 sec). A set of tumor-only projections (without thorax phantom in the beam path) were obtained and used to fit a sinusoid curve to obtain ground truth for the tumor motion.

2.C.3.b. Tumor tracking

A simple template matching method was used for tumor tracking. The tumor-only projections were used to determine the template. The template had a banded shape around the projected tumor edge, with a positive (+1) band inside the tumor edge, a negative (−1) band outside the tumor edge, and zero elsewhere, resulting in an average template value of 0. For each projection, the normalized cross-correlation (NCC) between the template and a 40 × 40 mm² ROI was calculated. The location with the peak NCC value was considered as the predicted location. The purpose of such a design is to maximize the NCC value from the tumor while minimizing the dependence on local signal intensity. We found a template with band thickness of 6 pixels provided the best performance for both SE and DE tracking tasks, so it was used for generating the following results. All the analysis was implemented using MATLAB R2016a (MathWorks, Natick, MA, USA).

2.C.3.c. Metrics

Tumor tracking performance was compared between SE and DE imaging techniques. The SE data used the projections obtained from the top detector layer. Tracking accuracy was estimated using the root mean square error (RMSE) between the predicted and ground truth tumor locations.

2.D. 3D imaging applications

DECT is able to perform quantitative material decomposition, including quantification of contrast agents, electron density, and effective atomic number and generating virtual monoenergetic images. In this section, a multi-energy CT phantom (Gammex, Middleton, WI) was used to test the DL FPD’s ability to perform quantitative material decomposition. The phantom was a water-equivalent, 20 cm diameter cylinder and contained several 2.85 cm diameter contrast inserts, including two calcium inserts (300, 100 mg/ml Ca), 3 iodine inserts (10, 5, 2 mg/ml I), 1 adipose insert, and 2 blood/iodine mixtures (4, 2 mg/ml I). A CBCT scan of the phantom was acquired with pulsed exposures of 120 kV, 20 mA, and 18 ms, synchronized to the detector. The detector was operated at 10 fps with 2×2 binning mode, and 500 projections were acquired per rotation. The same SAD and SID as in 2.C were used. For this part, we focused on demonstrating the detector’s potential to perform material decomposition without the influence of other artifacts. Thus, the vertical collimation was reduced to 4 cm at the detector to reduce scatter contamination. The projections were reconstructed to an isotropic voxel size of 0.5 mm with a standard filtered-backprojection (FBP) algorithm using Hanning apodization.

2.D.1. Material decomposition

We used an empirical dual energy calibration method^[31] to perform the material decomposition. This technique does not require knowledge of the imaging spectrum, the attenuation coefficients, or the detector response. As shown in Fig. 4, the decomposition was done in projection domain, where the basis material line integrals l₁ and l₂ (e.g., water and bone) are obtained as a polynomial function of the two sets of measured line integrals, l_L and l_H, respectively. The l₁ and l₂ are then reconstructed to obtain images corresponding to the basis materials.

Fig. 4.

Workflow of material decomposition.

In this work, we considered two sets of basis material pairs, i.e., water/calcium (l_W/l_Ca) and water/iodine (l_W/l_I). The coefficients {a} of the polynomial function were determined using a least-squares fit of the reconstructed polynomial line integrals to regions of known materials in the multi-energy CT phantom. For example, the line integral of water, l_W, was expressed as:

$$l_W=a_1{l}_L+a_2{l}_H+a_3{l}_L^2+a_4{l}_L{l}_H+a_5{l}_H^2+\cdots .$$ (4)

We found that a 4^th-order polynomial function was sufficient to perform the material decomposition and hence was used in this work.

2.D.2. Virtual monoenergetic images

The virtual monoenergetic (VM) images were synthesized as if the images were acquired with a monoenergetic x-ray beam. They were generated by weighting the mass densities of the two basis images by their mass attenuation coefficients at select energies, expressed as:

$$\mu \left(E\right)= {\left(\frac{\mu }{\rho }\right)}_1\left(E\right)\cdot {\rho }_1+ {\left(\frac{\mu }{\rho }\right)}_2\left(E\right)\cdot {\rho }_2$$ (5)

where ${\left(\frac{\mu }{\rho }\right)}_1\left(E\right)$ and ${\left(\frac{\mu }{\rho }\right)}_2\left(E\right)$ represent the mass attenuation coefficients of the two basis materials, and ρ₁ and ρ₂ are the mass densities of the basis materials. Theoretically, VM images should be free of beam hardening artifacts that are associated with polyenergetic spectra. To determine the optimal energy for beam hardening reduction, we generated VM images from 40 to 100 keV and selected the one with highest background uniformity. We also determined the optimal VM energy for maximizing the contrast-to-noise ratio (CNR) of the 10 mg/ml I and 300 mg/ml Ca inserts with respect to background water.

3. Results

3.A. Detector Performance

3.A.1. Pre-sampling MTF

Fig. 5 shows the pre-sampling MTF for the top and bottom layers of the DL FPD. Despite large differences between the RQA5 and 120 kV spectra, there was little impact on MTF. The top layer has a f_50% (the spatial frequency at which the MTF drops to 50%) of about 2.2 mm⁻¹, while the bottom layer is substantially lower at about 1.2 mm⁻¹. These differences are largely driven by the different CsI:Tl scintillator thicknesses, where the thinner top scintillator (200 μm) offers much higher spatial resolution than the thicker bottom scintillator (550 μm).

Fig. 5.

Pre-sampling MTFs for top and bottom layers measured with RQA5 and 120 kV spectra.

3.A.2. Normalized NPS

The detector NNPS are shown in Fig. 6 for the RQA5 and 120 kV spectra. The electronic noise in dark-field images was found to be substantially lower than noise in the flat-field images and not likely to influence the NNPS (standard deviation of 8.6 ADU in dark-field images for both layers vs >84.8 ADU in flat-field images for both spectra and both layers). Using a conversion factor of 487 e⁻/ADU for gain 5^[32], the electronic noise was 3900 e⁻. By coincidence, the top layer NNPS are almost identical for both spectra. The bottom layer shows an approximate 6-fold and 3.5-fold higher NNPS than the top layer at lower frequencies (<0.5 mm⁻¹) for the RQA5 and 120 kV spectra, respectively, due to absorption in the top layer and the 1 mm Cu filter. The greater increase in NNPS for RQA5 compared with 120 kV is due to its lower energy and therefore higher attenuation in the scintillator.

Fig. 6.

Detector NNPS for top and bottom layers measured with RQA5 and 120 kV spectra.

3.A.3. Detective quantum efficiency

Figure 7 shows the DQE calculated for both layers with RQA5 and 120 kV spectra. The $\frac{\varphi }{{ K}_a}$ for the RQA5 and 120 kV spectra are 263400 photons/mm²/mR and 208550 photons/mm²/mR, respectively, and the measured dose at the detector surface was 3.78 mR and 5.09 mR, respectively. The top layer has a fairly high DQE(0) of 42%, although, due to its thinner scintillator, is lower than the single-layer XRD 4343RF, which has a DQE(0) of 76% for RQA5^[33]. The top layer DQE is slightly higher for RQA5 than 120 kV, due to its higher efficiency at lower energies. The bottom layer has substantially lower DQE than the top layer due to its lower MTF and much higher NNPS (DQE (0) of 7% and 10% for RQA5 and 120 kV, respectively). The bottom layer DQE is higher for 120 kV than RQA5 due to the higher energies better penetrating through to the bottom.

Fig. 7.

DQE for top and bottom layers measured with RQA5 and 120 kV spectra.

3.B. 2D imaging applications

3.B.1. Dual energy radiography

Figure 8 shows the LE and HE projection images after log normalization and the impact of scatter on image quality. The contrast for both LE and HE projections is substantially improved after scatter correction. Figure 9 shows the material-specific projections, after weighted subtraction of the LE and HE projections. Note that the coronary artery contrast should have a negative soft tissue component and positive bone component since it is higher Z_eff than both soft tissue and bone. Scatter contamination reduces the accuracy of the line integrals and therefore biases the material images. As indicated by the red arrows, without scatter correction, the contrast of barium is inverted in the soft tissue and bone images, and soft tissue and bone are not separated well. These errors are resolved when using the scatter-corrected LE and HE images to generate the material-specific projections. Note that barium in the coronary arteries appears in the bone image, which is a known limitation of DE imaging. The soft tissue image also contains empty voids after bone removal, such as in regions of the vertebrae. This is a combined effect of reduced soft tissue path length after bone removal and air within the vertebrae of the phantom.

Fig. 8.

Impact of scatter on LE and HE projections. The top row shows the LE and HE projection images without scatter correction, while the bottom row is with scatter correction. Display window: [1 6] (unitless).

Fig. 9.

Material-specific projections of soft tissue (left) and bone (right). The top row is without scatter correction, while the bottom row has scatter correction. Red arrows indicate the regions with improved material classification after scatter correction.

3.B.2. Tumor tracking using DE fluoroscopy

Figure 10 depicts the tumor tracking results for SE and DE imaging over two respiratory cycles (10 sec). DE imaging substantially improved the tracking accuracy compared to SE. The RMSE for SE and DE tracking results compared to the sinusoid fit of ground truth locations are 2.63 mm and 0.25 mm, respectively. SE tracking usually failed when the tumor edge overlapped the ribs. Figure 11 shows an example of this scenario. Even with the outliers near Y = −5 mm removed, the SE RMSE was 0.33 mm. Note that some residual bone structures are seen in the soft tissue image, which could be a limitation of the weighted subtraction approach, as well as different spatial resolution between the layers.

Fig. 10.

Comparison of tumor tracking results using SE and DE imaging to ground truth locations.

Fig. 11.

Zoomed-in views of an example where the tumor edge is overlapped with ribs. The template is derived from the ground truth image, which shows a scale of 1 cm at the detector. The tracked position for each image is indicated by the cross marker, where the soft tissue image generated using DE imaging better resolves the tumor edge and predicts the location.

3.C. 3D imaging application: Quantitative Cone-beam CT

3.C.1. Material decomposition

Figure 12 shows an axial view of the reconstructed LE and HE images and the results of material decomposition for water/calcium and water/iodine basis material pairs. As summarized in Table 2, the estimated material densities are 294.68 ± 17.41 mg/ml and 92.14 ± 15.61 mg/ml for 300 mg/ml and 100 mg/ml calcium inserts, respectively, and 8.93 ± 1.45 mg/ml, 4.72 ± 1.44 mg/ml, and 2.11 ± 1.32 mg/ml for 10 mg/ml, 5 mg/ml, and 2 mg/ml iodine inserts, respectively, resulting in an average error of < 5% compared to ground truth. Some artifact is visible in the material decomposition images, possibly due to residual scatter or spectral non-uniformity in the x-ray beam or detector response.

Fig. 12.

Material decomposition results with DL FPD. Top row: DE CBCT images, window: [–330 560] HU. The middle and bottom rows show the material decompositions for water/calcium and water/iodine, respectively.

Table 2.

Quantitative summary of estimated material densities

Water/Calcium	Estimated Material Densities (mean ± std, mg/ml)
Water (1000 mg/ml)	990.37 ± 48.98
Calcium (300 mg/ml)	294.68 ± 17.41
Calcium (100 mg/ml)	92.14 ± 15.61
Water/Iodine
Water (1000 mg/ml)	986.51 ± 45.65
Iodine (10 mg/ml)	8.93 ± 1.45
Iodine (5 mg/ml)	4.72 ± 1.44
Iodine (2 mg/ml)	2.11 ± 1.32

3.C.2. Virtual monoenergetic images

The synthesized VM images from 40 keV to 100 keV are shown in Fig. 13 (left panel). In principle, the monoenergetic images should be free of beam hardening artifacts for the entire energy range since they are a linear weighted combination of basis material images. However, in practice it is difficult to achieve perfect VM images due to material decomposition error from imperfect dual energy calibration and residual scatter. With the 120 kV spectrum and prototype DL FPD, we found that the VM image at 60 keV achieved the highest image uniformity and contained the least beam hardening artifacts. Fig. 13 (right panel) shows the comparison of estimated and theoretical mono-energetic HU values for 10 mg/ml iodine and 300 mg/ml calcium. It is seen that the estimated HU values accurately predicted the ground truth, with small errors at 60 keV for both inserts (<2%). CNR for the 10 mg/ml iodine and 300 mg/ml calcium inserts as a function of VM energy are plotted in Fig. 14. Similarly, the maximum CNR is achieved at around 60 keV for both inserts. CNR rapidly drops for energies away from the optimal energy, with increased noise at lower energies (despite higher contrast) and with decreased contrast and increased noise at higher energies, as expected^[34].

Fig. 13.

Left: VM images from 40 keV to 100 keV. The 60 keV image (bounded in red) provides optimal image uniformity and has maximum CNR for both iodine and calcium inserts. Window: [–330 560] HU. Right: The comparison of estimated and theoretical HU values for 10 mg/ml iodine and 300 mg/ml calcium inserts.

Fig. 14.

CNR for 10 mg/ml iodine and 300 mg/ml calcium as a function of VM energy.

4. Discussion and Conclusion

In this paper, we characterized a prototype DL FPD and demonstrated its ability to perform single-exposure DE radiography/fluoroscopy and DE-CBCT with accurate quantification of iodine and calcium concentrations. The advantages of dual energy imaging with a dual layer detector over other methods are 1) superior spatial and temporal registration between the constituent LE and HE images, and 2) ease of implementation since special imaging protocols (e.g., kV switching) are not required. Weaknesses include its reduced overall DQE and the reduced spectral separation of the constituent images.

Detector performance characterization included measuring spatial resolution, image noise, and DQE. Results showed the MTF of the (thinner) top layer was higher than that of the bottom layer since spatial resolution is mainly determined by the scintillator thickness. A significant increase in NNPS and reduction in DQE were observed for the bottom layer, mainly due to photon loss in the top layer and the 1 mm Cu filter that was in place to increase energy separation. Because both high energy separation and detector DQE are important for obtaining good material decomposition and overall image quality, in the future we will seek to optimize detector performance by studying the tradeoffs involved in using different combinations of scintillator and filter thicknesses and materials. Additionally, such DQE characterization can help inform us of how to combine the images in an optimal way that takes advantage of their different spatial resolution and noise properties to maximize DQE as a function of spatial frequency^[35]. In addition, we will explore the optimal source spectrum (kV and filtration) that leverages the DL FPD for a given imaging task.

For 2D imaging applications, we have shown that the DL FPD can generate high-quality material-specific images. With the capability of operating at 15 fps in 2×2 binning mode, the DL FPD is suitable to many clinical tasks in image-guided radiotherapy and interventions while the alternative approaches such as fast kV switching typically halve the frame rate and are susceptible to motion artifacts, although it should be noted that registering the images can mitigate some of these artifacts^[13]. We are planning to conduct additional studies investigating the real-time tracking capability of this detector, including comparisons with methods such as fast kV switching.

Scatter contamination is known to be detrimental to image quality for single-energy radiographs and CBCT. For DE imaging, scatter also negatively impacts material decomposition accuracy. Conventional methods use an anti-scatter grid or bowtie filter to reduce scatter contamination for CBCT. Such methods are effective and robust; however, it brings additional considerations for dual energy imaging. Anti-scatter grids increase the beam filtration and may have non-uniform regions if not carefully aligned. Bowtie filters alter the dual energy spectra across the entire detector FOV, complicating the calibration process for material decomposition. Similarly, the heel effect may introduce spatial variation of the spectra and would require position-dependent spectral calibration. In this work, the scatter distributions for the two layers were estimated and subtracted from the raw projections. This scatter correction was instrumental in generating accurate material-specific images. Further improvements could be made by accounting for scatter within the detector (i.e., x-rays scattering between the layers), as this was not explicitly modeled. In the future, comprehensive Monte Carlo studies will be conducted to investigate the impact of scatter on image quality for different detector designs.

For 3D applications, we demonstrated that successful basis material decompositions can be obtained for DE CBCT. The material-specific CBCT images were then used to generate VM images. An optimal VM energy at 60 keV was observed to achieve maximum image uniformity and CNR of iodine and calcium inserts. The ability to generate high-quality VM images across a wider energy range would be improved with increased DE spectral separation as well as increased detector DQE. These tasks will be a focus in future DL designs and optimization studies. As with 2D applications, a thorough comparison between the DL and other DE methods such as kV switching will be conducted in subsequent work to quantify the advantages and disadvantages of the different approaches for 3D applications. With material decomposition tasks as a focus in this work, the 3D images were reconstructed to relatively large voxel sizes. In the future, we will explore clinical applications that could take full advantage of the high spatial resolution of this detector, such as angiographic imaging, bone imaging, etc.

In summary, we have evaluated a prototype DL FPD in several DE applications. It is capable of capturing DE images in a single exposure, giving it superior spatial and temporal registration for fluoroscopy/radiography applications where motion may be a concern. Additionally, the DL FPD provides DE-CBCT in a single rotational acquisition, without having to switch energies or perform a second acquisition. These early results are promising, and future work will seek to optimize the detector and evaluate it in clinical applications.