Tomographic Cherenkov-excited luminescence scanned imaging with multiple pinhole beams recovered via back-projection reconstruction

Abstract

Cherenkov-excited luminescence scanned imaging (CELSI) is achieved with a clinical linear accelerator during External Beam Radiotherapy, to map out molecular luminescence intensity or lifetime in tissue. In order to realize a deeper imaging depth with a reasonable spatial resolution in CELSI, we optimized the original scanning procedure to complete this in a similar way to x-ray computed tomography and with image reconstruction from Maximum-likelihood expectation maximization and multi-pinhole irradiation for parallelization. Resolution phantom studies showed that a 0.3 mm diameter capillary tube containing 0.01 nM luminescent nanospheres could be recognized at a depth of 21 mm into tissue-like media. Small animal imaging with a 1 mm diameter cylindrical target demonstrated that fast 3D data acquisition can be achieved by this multi-pinhole collimator approach to image high resolution luminescence through a whole animal.

Cherenkov-excited luminescence scanned imaging (CELSI) uses the collimating system of radiotherapy linear accelerator (LINAC) to send a sheet of radiation travelling across the imaged subject in a manner that is equivalent to the excitation-beam shaping used in light-sheet microscopy [1,2]. In CELSI, luminescence and Cherenkov image can be acquired in the same view for molecular and anatomical imaging, respectively, but with different pulse delay and integration time. By restricting the X-ray excitation to a single, narrow sheet, the origin of the optical photons can be inferred regardless of where these photons were detected, and how many times they scattered in tissue. Depth information for estimating intensity attenuation and emission diffuse can be obtained according to the distance from the medium surface, which can be calculated from the recovered Cherenkov images. Previous demonstrations of CELSI showed that superb spatial resolution on the order of 100 microns can be achieved within a depth of 5 mm. In this letter, a novel image modality named tomographic CELSI (T-CELSI) was proposed to take advantage of the CELSI as combined with computed tomography algorithms.

X-ray luminescence computed tomography (XLCT) has been demonstrated with tomographic X-rays stimulating luminescence emissions with linear reconstruction [3], and uses continuous diagnostic kV energy X-ray photons, below the Cherenkov threshold, allowing for direct excitation of nanoparticulate contrast agents via the photoelectric effect and optical emission from radioluminescence. In XLCT, tomographic data is generated by irradiating the object in a CT scanning method, where a sequence of programmed X-ray beams are systematically projected at each position and beam angle. Since measured photons were created somewhere on the narrow path of the X-ray beam along different beam directions, the optical sensor is not required to spatially resolve photons. Instead, all the pixels can be summed up to obtain a higher signal-to-noise ratio (SNR), as has been demonstrated in CELSI. By arranging an X-ray detector at a position orthogonal to the optical detector, anatomical imaging is performed simultaneously with molecular imaging via standard X-ray computed tomography (CT). Then the co-registered anatomical images with extra auxiliary like fiducial points can be used to extract depth information for the tomographic algorithm.

In the proposed T-CELSI method demonstrated here, the design to optimize this was: (1) a single-pixel camera concept was introduced to increase imaging depth via better integrated signal strength; (2) boundary information for estimating target depth was obtained from Cherenkov images at each incident beam location, which could perfectly match the time-delayed luminescence data without any need for co-registration; and (3) based on a standard LINAC, the X-ray beam sequence and multiple projections were realized with multi leaf collimator (MLC) and gantry movement, respectively.

In T-CELSI imaging, selective pulsed X-ray beams were emitted from a LINAC (Varian Clinac 2100 CD, Varian Medical Systems, USA), with parameter settings as follows: photon energy of 6MV, dose rate of 600 monitor units per minute, and a fixed repetition rate of 360 Hz. As shown in Fig. 1, 1×1 mm² square beam was raster scanned with the MLC, and the emitted optical photons on the top surface were imaged by an intensified time-gated camera (C-Dose, Dose Optics, Lebanon New Hampshire) worked with a focal length of 50mm and f/1.2 lens. The ICCD was time-gated for Cherenkov acquisition with no pulse delay and a 3.5μs integration time to capture just during the LINAC pulse, and then for luminescence detection with a 4.26μs delay from the LINAC pulse and 100μs integration time. Of note, the camera worked in a single-pixel mode. Europium chelate microspheres (Bangs Laboratories, Inc.) were used as nanophosphor, which had 0.3μm diameter, broadband UV excitation and peaked 605–625nm emission, and approximately 2000μs lifetime. These particles can be made biocompatible with coating and ultimately functionalized with binding proteins for molecular imaging applications. Half-angle projection images were obtained with θ ∈ [0°,180°] and a step of 5°, i.e., 37 angular positions. The acquisition time for each projection was 20 seconds for a total approximate dose of 2Gy. Squared beams were continuously scanned without stops to save time, and 135ms was assumed to as an exposure time leading a total scanning/radial positions of 150. All the phantoms or the mouse were placed at the LINAC isocenter.

Fig. 1.

(a) Measurement geometry for tomographic CELSI. At each gantry position, scanning beams are generated by the programmable MLC (b).

In the T-CELSI system, X-rays travel through the tissues with the expected secondary electron build up over the first 1.5cm typically for a beam energy of 6MV [4], followed by decay for deeper depths, and this electron dose induces Cherenkov light that secondly excites the nanophosphors. Image reconstructions were twofold involved: (1) a standard maximum-likelihood expectation-maximization (ML-EM) method in X-ray CT was used to recover Cherenkov light distribution as it is seldom scattered [5]; and (2) a modified ML-EM (mML-EM) was designed to solve luminescent concentration ρ(r) at position r. Here the k-th (k≥0) updated ρ^(k)(r) on Y-Z slice [i.e., r=(y, z)] is given as follow:

$${\rho }^{\left(k+1\right)}\left(r\right)=\frac{{\rho }^k(r)}{\mathcal{R}*\left[\overline{\Gamma }\right]}\mathcal{R}*\left[\frac{{\Gamma }_{\text{meas}.}\left(y,\theta \right)}{{\Gamma }^{\left(k\right)}\left(y,\theta \right)}\right],$$ (1)

where $\mathcal{R}*$ [∙] represents the inverse Radon transform; Γ_meas. or Γ is the emission light intensity measured or predicted at the top surface; and $\stackrel{-}{\mathrm{\Gamma }}$ is an all-ones vector with the same as Γ or Γ_meas., which is used to normalize ρ^(k)(r). Here ρ⁽⁰⁾(r) is assumed to be an all-ones matrix. Note that the Γ(y,θ) image is actually a sinogram as exampled in Fig. 2(c), where the number of y positions are the same as that of the MLC scanned squares. Eq. (1) can be interpreted with the ML-EM principle as [5]: the backprojection of the ratio between Γ_meas. and Γ is used as a multiplicative coefficient to update ρ(r). Here the forward solution of emission light intensity Γ measured at the top surface under a specified gantry angle is calculated by:

$$\{\begin{array}{l}\text{P}\left(z,\theta \right)=e^{-{\mu }_{eff}\left(z-z_{\text{max}}\right)}/{\left(f\left(\theta \right)+z\right)}^2\hfill \\ Q^{\left(k\right)}\left(r,\theta \right)=\eta \text{P}\left(z,\theta \right){\rho }^{\left(k\right)}\left(r\right)\hfill \\ {\Gamma }^{\left(k\right)}\left(y,\theta \right)=Q^{\left(k\right)}\left(r,\theta \right)\otimes \text{G}\left(r,y\right)\hfill \end{array},$$ (2)

where f is the source to surface distance (SSD); z_max represents the depth of the maximum value of relative absorbed dose P [6]; μ_eff is the effective X-ray linear attenuation coefficient for the primary beam inside the medium; Q(r) is the source term; η is the emission quantum yield ratio between X-ray and luminescence light; ⊗ is the convolution operator; G(r,ԑ) are the system Green functions for luminescence light transport, which was calculated for diffusion equation combined with a Robin-type boundary condition [7]. Any other modeling for photon transport in tissue could be used to solve Q(r) ⊗ G(r,ε), e.g., the Monte-Carlo method. Since the Cherenkov and luminescence images were detected by a same camera just with different time delays, these two images can be co-registered well and the Cherenkov-extracted surface was used to estimate the z-distance from the reconstructed target.

Fig. 2.

Resolution phantom imaging. (a) Y-Z view of a cylindrical tank as shown in Fig. 1(a) incorporated with capillaries of different diameters and overlaid with a recovered Cherenkov image. (b) X-Y view of the capillaries photograph, merged with a Cherenkov image to indicate the X-ray beam region. (c) Exampled raw images at several radial positions with 0° projection, and the raw sinogram. (d) Recovered targets and associated sinograms via sIRT and mML-EM algorithms. (e) Y-profiles plotted along the dashed white lines in (d), and (f) contrast to background ratios for different capillary sizes and target depths.

A resolution phantom was designed to investigate the influence of lesion size on detectability. As shown in Fig. 2(a), four cylindrical capillaries, with diameter 0.3, 0.5, 0.7, and 0.9 mm, were filled with phosphor concentrations of 0.01 nM and embedded at depths of 0.3, 1.2, and 2.1cm inside the tissue-mimicking phantom, which was %1 porcine blood mixed (Lamphire Inc, Pipersville PA) with 1% Intralipid (dilute from 10% Intralipid, Sigma-Aldrich). As a validation of using Cherenkov signal to locate medium boundary, the reconstructed Cherenkov Y-Z slice was overlaid in Fig. 2(a). An X-Y view of Cherenkov MIP image at 90° beam projection (i.e., when the collimator was opposite to the detector) is shown in Fig. 2(b), merged with unfilled phantom under roomlight. Caused by diffusion and divergent X-ray irradiation from the LINAC collimator, the beam band covering 90% intensity maximum had a width of 8mm, much larger than the original 1mm. The raw image captured for 4 radial positions at 0° projection is shown in Fig.2 (c), as well as the raw sinusoidal trajectories of the twelve cylinders. Based on the raw sinogram, reconstructions via a standard inverse Radon transform method (sIRT) and the mML-EM with 50 iterations are compared, and the recovered ρ(r) and Γ(y,θ) are shown in the upper and bottom line of Fig. 2(d), respectively. For those 0.3-mm diameter capillaries, Y-profiles along the dashed white lines in (d) are shown in Fig. 2(e). Signal-to-noise ratio (SNR) for all the recovered 12-targets via the sIRT and mML-EM methods are plotted in Fig. 3(f).

Fig. 3.

A sensitivity phantom imaging: (a) geometry sketch of the phantoms, (b) reconstructed results via sIRT and mML-EM methods, (c) resultant contrast-to-background ratio curves, and (d) averaged SNR versus concentrations of Europium.

A sensitivity phantom was imaged to assess the minimum detectable nanophosphor distribution for different target depths. Three 0.7-mm-diameter cylinders were filled with nanophosphor concentrations of 0.005, 0.01, and 0.05 nM and embedded at depths of 1, 2, and 3 cm, as shown in Fig. 3(a). To further improve the data acquisition efficiency, half of the original projections (i.e., 19 angles and θ ∈ [0°,90°]) were adopted to generate the raw sinogram as a proof-of-concept. Reconstructed results are present in Fig. 3(b), and the calculated SNR against target depth are depicted in Fig. 3(c). Then there SNRs were averaged ($\overline{\text{SNR}}$) over different depths, and their relation to concentration is plotted in Fig. 3(d).

For in-vivo imaging, all animal procedures were approved by the Dartmouth Institutional Animal Care and Use Committee, and the studies here were carried out in compliance with these approved procedures. Nude female mice were purchased at 6 weeks of age from Charles River Labs. To evaluate the performance of the proposed method, we implanted a transparent tube (inner diameter 1.0mm) at a depth ranging from 1cm to 2cm, which was filled with the nanophosphors at a concentration of 1.1nM, into the body of mouse near the liver inside the peritoneal cavity.

In order to 3D render the luminescent target, a multi-pinhole collimator was designed to provide the complete data necessary as exampled in Fig. 4(a). Different from the multi-pinhole XLCT[8,9,10], no extra beam shaping appliance is requisite. The measurement geometry is shown in Fig. 1 with 37 angular positions. By making use of the MLC control, an array of multiple pinholes could be moved vertically in parallel. For this 3D scenario, there were a cluster of raw sinograms depending on the number of pinholes used, and each pixel was the sum of several lined values (depending on the size of the squared beam) of the raw image. 3D image was achieved by stacking all the 2D reconstructed results along X-direction. Figure 4(b) shows the recovered cross-section of merged luminescence, Cherenkov, and a CT image, which was obtained from a commercial imager (IVIS SpectrumCT, PerkinElmer, USA) with source voltage 50kVp and current of 1mA. We can see that, the glass-made tube recovered with T-CELSI shows good agreement with that from CT. The tube could be found with an ellipse shape due to the off-axis placement, and some white areas inside the body were thought to be air in the digestive tract. The luminescence maximal intensity projection (MIP) on X-Y plane is shown in upper of Fig. 4(c), where the dashed box indicates the original position and profiles along the colored dashed lines in (c) are plotted in the bottom with full width half maximum (FWHMs) of 1.0mm (blue), 1.1cm(red), and 1.3cm(yellow). Differences among these profiles was caused by the different imaging depths: 1cm(blue), 1.5cm(red), and 2cm(yellow) as indicated in Fig. 4(d).

Fig. 4.

3D in-vivo T-CELSI imaging demonstration of a 1mm-diameter tube loaded with nanophosphors: (a) detailed multi-hole irradiation design from the linac MLC, (b) Cherenkov, luminescence, and CT overlay axial images sliced along the dashed black line in (d), (c) luminescence MIP image on the X-Y plane, and (d) merged luminescence and CT images for sagittal and coronal planes.

Similar to CELSI, the whole measurement geometry in T-CELSI relies on a radiotherapy LINAC, which could provide a pathway for specified molecular sensing with the T-CELSI during external beam radiation therapy, since sophisticated dynamic trajectories involving gantry and MLC motions could be always found in most modern radiation treatment plans, e.g., Intensity Modulated Radiation Therapy (IMRT), and Stereotactic Body Radiation Therapy (SBRT). However, T-CELSI maintains higher resolution and molecular sensitivity at significant depths inside tissue. As exampled in Fig. 3(c), only the superficial targets can be clearly recognized as numbered with Position 80. As a result, stacking these noisy images for a 3D rendering as in CELSI could be challenging. In comparison, each pixel in the sinogram of Fig. 3(c) is a sum of the corresponding raw image, which greatly enhances the measurement SNR in T-CELSI. Furthermore, the raw sinogram was smoothed along each column to perform a stable reconstruction. Other specific denoising strategies on low-dose CT could also be employed [11,12].

As in most XLCT modalities, Cherenkov images were instead used to provide anatomy information for image reconstruction. We can see from Fig. 2(a) and Fig. 4(b) that the outline of Cherenkov distribution matches well with the target surface towards to the camera. Comparing to the CT-based prior information in XLCT, Cherenkov images can be better matched with luminescence ones without delicate co-registration. Limited by a fixed view ranging to 180° with one mirror, some surface places might not be visible in extreme cases, e.g., the portions being much closer to the bench as circled in Fig. 4(b). For those cases, a smooth interpolated connection for a complete boundary could be reasonable for most tissue applications. An alternative is to use sufficient mirrors or cameras.

Because the Cherenkov and luminescence beam are both attenuated and diffused while traversing the sample, the sinogram should not perfectly symmetrical, which is the reason why it is preferable to scan the phantom over the full 360° range. For practical purpose, acquisition time or radiation dose should be optimized. As a preliminary trial, reconstruction based on half (180°) and quarter (90°) projection range were discussed. The reconstruction performance clearly depended upon the angular positions used, and so only the most promising choices of angles were shown here. In addition, a full 3D visualization was demonstrated with a multi-pinhole illumination to achieve the same detection efficiency as a 2D process as shown in Fig. 4(a). To save Z-resolution, the gap distance d was set to 1cm to circumvent cross-talk effect between beams. However, limited by the mechanical size (1mm width for central 32 pairs of leaves) of the MLC leaf, the X-resolution was inherently lower than the Y- or Z- direction resolution values, depending on the squared scanning interval. By taking advantage of flexible patterning with MLC, further accelerate acquisition could be achieved by irradiating multiple point lines for 3D imaging in a way similar to the coded CELSI [13] Other algorithms fairly developed for sparse view XLCT or CT could be borrowed in T-CELSI [14–16].

In comparison to traditional fluorescence molecular tomography (FMT) [17], fluorophore positions could be always accurately recovered in T-CELSI. As exampled in Fig. 2 (d), the geometric centers of the targets can be found to agree well with the real ones regardless of the reconstruction methods. This is attributed to the narrow X-ray beam, and thus the target can be localized accurately with a back-projection algorithm. By bringing optical tomography technique with the ML-EM framework, reconstructions of target intensity can basically get rid of depth dependent (reconstructed signal via mML-EM varied by <3% when the depth changed from 1 to 3 cm), which was quantized with signal-to-noise ratio (SNR) as plotted in Fig. 3(e) and Fig. 4(c). After correction, an approximated linear relationship can be found in between $\overline{\text{SNR}}$ and nanophosphor concentration.