Multi-beam scan analysis with a clinical LINAC for high resolution Cherenkov-excited molecular luminescence imaging in tissue

Cherenkov-excited luminescence scanned imaging (CELSI) is achieved with external beam radiotherapy to map out molecular luminescence intensity or lifetime in tissue. Just as in fluorescence microscopy, the choice of excitation geometry can affect the imaging time, spatial resolution and contrast recovered. In this study, the use of spatially patterned illumination was systematically studied comparing scan shapes, starting with line scan and block patterns and increasing from single beams to multiple parallel beams and then to clinically used treatment plans for radiation therapy. The image recovery was improved by a spatial-temporal modulation-demodulation method, which used the ability to capture simultaneous images of the excitation Cherenkov beam shape to deconvolve the CELSI images. Experimental studies used the multi-leaf collimator on a clinical linear accelerator (LINAC) to create the scanning patterns, and image resolution and contrast recovery were tested at different depths of tissue phantom material. As hypothesized, the smallest illumination squares achieved optimal resolution, but at the cost of lower signal and slower imaging time. Having larger excitation blocks provided superior signal but at the cost of increased radiation dose and lower resolution. Increasing the scan beams to multiple block patterns improved the performance in terms of image fidelity, lower radiation dose and faster acquisition. The spatial resolution was mostly dependent upon pixel area with an optimized side length near 38mm and a beam scan pitch of P = 0.33, and the achievable imaging depth was increased from 14mm to 18mm with sufficient resolving power for 1mm sized test objects. As a proof-of-concept, in-vivo tumor mouse imaging was performed to show 3D rendering and quantification of tissue pO2 with values of 5.6mmHg in a tumor and 77mmHg in normal tissue.


Introduction
Radiation therapy induces Cherenkov light emission within tissue, and this signal provides potential for direct molecular sampling of the tissue microenvironment [1][2][3][4]. The approach to using Cherenkov Excited Luminescence Scanned Imaging (CELSI) was examined here, specifically varying the spatial patterns used to excite Cherenkov light from megavoltage (MV) X-rays in external beam radiation therapy (EBRT) [5,6]. CELSI can provide threedimensional luminescence images through layer-scanned Cherenkov light sheet, which is carried out by movements of the multi leaf collimator (MLC) on the linear accelerator (LINAC), which is commonly used to provide conformal shaping of radiotherapy treatment beams [7]. This application has been experimentally demonstrated in-vivo and is hopeful for clinical trial use, taking advantage of the dynamics already inherent in radiotherapy treatments [5,6]. In this study, progressive approaches to scan beam size and shape were explored to understand how incorporation with dynamic treatment plans might occur.
Since tissue partial pressure of oxygen (pO 2 ) is well known to affect the efficacy of radiotherapy [1, 8,9], imaging this signal has been a primary focus of previous CELSI studies. In order to image pO 2 in tissue for EBRT, megavoltage (MV) photon beams have been utilized to induce Cherenkov emission in thin sheets within the tissue which in turn excites an oxygen-sensitive phosphorescent probe. An example is excitation of Platinum (II)-G4 (PtG4) [10][11][12] via Cherenkov light emission, which efficiently absorbs blue light (λ max = 435nm) and emits phosphorescence in the NIR region (λ max = 772nm). Epi-illumination based CELSI in tissue allows for imaging to a depth of one centimeter or more, depending on the fluorophore concentration and radiation dose used. The scattering-induced blurring can be compensated by deconvolution processing based on the point spread function (PSF), evaluated by radiative transport models [6,13,14]. The hypothesis examined in this study was that when designing this irradiation scheme, the beam shape could be chosen to optimize the spatial-resolution of recovered CELSI images. Further, it might be possible that existing clinical treatment plans allow for high resolution imaging.
Patterned illumination schemes have been investigated more recently with the emergence of compressed sensing, where computed image recovery is utilized to improve the image quality without extra radiation exposure and/or acquisition time [15]. While the ultimate extent of compressed sensing methods uses complex patterns, the methodology for this could include a simpler spatial-frequency-domain imaging (SFDI) approach or more complex techniques such as used in coded aperture imaging (CAI). Current SFDI instruments typically use projected spatially modulated light, projected onto the sample, and then knowledge of the illumination patterns are used to demodulate the images [16][17][18]. CAI requires a coded mask placed either, in front of the radiation beam source to shape the incident beam, and at the detector to block the transmitted beam and disambiguate scatter angles [19][20][21]. For both of these structured imaging modalities, the excitation beam is modulated by a projector or a coded aperture, and the measured reflectance or transmittance is demodulated in postprocessing. This modulation-demodulation method (MDM) has the capacity to mathematically filter the low frequency out-of-focus scattered photons beyond the depth of field of the imaging objective. For the MDM process using a point raster scanned beam, the narrow beam acts as a virtual pinhole, analogous to the emission light captured through a pinhole in confocal microscopy. Diversification of the illumination pattern based on an MLC is clinically convenient in CELSI and could be time-saving, and in principle the type of spatially modulated patterns could be extremely beneficial to provide high resolution imaging. In particular, using coded apertures in conjunction with multiplexed illumination can reduce the acquisition time, given that projections from more than one source may be acquired in a single step [19].
Motivated by the idea of using patterned illumination schemes, the study here examined if the spatial resolution and contrast recovery of CELSI could be improved on using different spatial illumination patterns and post processing with MDM. The input radiation beam was modulated and raster-scanned by programming the MLC, and post-processing demodulation was performed by deconvolving the measured luminescence image with the kernel function of the Cherenkov image. This approach, referred to as MDM-CELSI, examined several EBRT-compatible MLC plans, and specifically how the measurement geometry affected the acquisition time, and radiation does. Modulation of the radiation (X-ray) beam was investigated using basic illumination patterns including squares and lines, and multiple combinations of these. Fig. 1. The experimental setup used to investigate the effect of radiation beam size is shown in (a) with a photo of the experimental setup with LINAC, phantom and ICCD camera; (b) the internal phantom geometry; (c) the basic N × N square-raster scanning patterns tested initially with different pixel areas and the luminescence images measured for N = 2 (below this), and (d) the calculated Signal to Background ratio (SBR) versus target area for N = 2, 4, 6, and 8, respectively.

Incidence modulation
A clinical radiotherapy accelerator (Varian Clinac 2100 CD, Varian Medical Systems, USA) was used to generate a pulsed beam with the following parameters: photon beam energy of 6 MV, dose rate of 600 monitor units per minute (MU/min), at a fixed repetition rate of 360 Hz, and a ~3.25 μs pulse duration. The X-ray beam impinged on the tissue-equivalent phantom, embedded with a 70 µM PtG4 target (inner diameter Ø ≈3mm, and length of 30mm) at a depth of ~1.2cm, which is detailed in Fig. 1(a) and (b). The camera with a focal length of 85mm and f/1.2 lens was mounted close to LINAC as shown in Fig. 1(a), resulting into an inclined angle of ~30° and a distance of ~1,500 mm from the sample. The luminescence signal was detected with time-gated acquisition of the ICCD (100 μs integration time and 4.26 μs delay after each x-ray pulse). The luminescence images were acquired with 100 × gain on the ICCD intensifier, 30 pulses integrated on-chip as accumulations prior to each frame readout, and 2 × 2 hardware pixel binning at readout, from chip size of 1024 pixels × 1024 pixels. Cherenkov images were acquired in the same way, but with 0 μs pulse delay and a 3.5 μs integration time. Measurement at an increased pulse delay of ~500 μs was also acquired as 'background images' from the same field of view (FOV), which was used to subtract out stray signals in the camera and room.
In order to utilize LINAC-based EBRT for the CELSI illumination plan, a patterned beam design must be created with consideration of the geometry and orientation the MLC, which does put some limit onto what is possible. For the MLCs used here, the lateral length (along the direction orthogonal to movement) of a single MLC leaf was 5mm with step sizes of 0.1mm; the smallest point excitation available realistically was on that spatial scale. The geometry of these MLCs are optimal for targeting single-connected regions or distributions along the central axis, but it can be impossible to form more complicated shapes that might be optimal, like the Hadamard matrices for example [22]. X-ray beams were raster-scanned with different sized squares as shown in Fig. 1(c), which was performed by delivering a custom dynamic dose plan, created in MLC code [13]. The measured luminescence images for an N × N (N = 2) raster-scanned illumination are shown below Fig. 1(c), where the regions encircled with black lines are the pixel areas included within the radiation beam, named as the effective coverage. The relationship of Signal to Background ratio (SBR) with the effective coverage, and N, are shown in Fig. 1(d), where the SBR can be found strongly dependent upon both. A reasonable square/beam area could be more than 1mm 2 . To accelerate the acquisition process, several square-illuminations are simultaneously enabled, which allowed for reduction of total scan time. As a proof-of-concept, one type of multi-square illumination with equal spaces is shown in Fig. 2 (c), where the dashed squares represent the next scanning positions. The multi-square distribution depends on following two parameters: (1) the square side-length l, which is found by a compromise between spatial resolution and measurement SNR; and (2) the beam pitch ' / P d l = , where d' is the step size. P <1.0 denotes there is an overlap in between the continuous squares and vice versa. For fullfield illumination, we assume P = 1, which indicates a contiguous scanning for adjacent positions. For multiple scanning, , d min is near 2mm. As an alternative way to accelerate acquisition, line-scan field illumination has been demonstrated in confocal microscopy [23], optical coherence tomography (microscopy) [24,25], and laminar optical tomography [25]. Further accelerations can be similarly reached by multiple enablement strategy, named as multi-lines illumination, as shown in Fig. 2(d). Also, the distance in between each line should be greater than d SNR . Comparisons between point-and line-shaped patterns with different combinations of l and P was examined. The procedure of image demodulation is illustrated by the flow chart in Fig. 3. For simplicity, a 2D square-raster scan is illustrated here with 4 × 4 uniformly distributed scanning positions. The resulting input raw image is composed of 16 pixels corresponding to the 16 positions. The steps are: (1) the raw images were processed frame-by-frame to mainly extract the region of interest through subtracting 'background' ones; (2) the Cherenkov image was used as the kernel function to deconvolve the corresponding luminescence images; and (3) demodulation is done by combing all the frames into a Maximum Intensity Projection (MIP) image. Considering the multiple deconvolutions in step 2, an economical one-step deconvolution algorithm was used for ith scanning position, which can be formulated as

Image demodulation
where Y and X is the raw and undistorted luminescence image, respectively; C is the Cherenkov image; ԑ is an infinitesimal integer; and F ⋅ and 1 F − ⋅ represent Fourier and inverse Fourier transforms, respectively. Note that the range of spatial frequencies w can be physically limited in a radius of R ( [ ] 1.5 2.5 ∈ in this paper), which is regulated by a specified beam pattern, and can be estimated instead through the high signal-to-background ratio Cherenkov image. A series of phantom studies were carried out using a tissue-equivalent phantom, which consisted of 1% Intralipid and 1% porcine blood per unit volume of solids, with absorption and reduced scattering coefficients estimated to be  Fig. 1, and the phantom adopted (#2) is illustrated in Fig. 4, where two plastic capillaries (inner diameter Ø≈1mm) containing 70 μM PtG4 solution were placed with an angle of 5° in between as shown in Fig. 4. The angled capillaries were designed to pick up a reasonable Y-period between unresolvable and completely resolved status for X-profile analysis. A holder, which is fixed to the wall of phantom container, was used to clamp the capillaries. Without specific notification, the depth of target is 8mm.

Comparisons between line-and square scanning pattern
In the first study, line-and square-raster scanning were compared for P = 0.50, 0.33, and 0.25, respectively, and a fixed square size of l = 77 mm, as shown in Fig. 5. To quantify the spatial resolution performance, a metric of success was introduced, defined as R x-profile of the image, essentially providing a relative measure of contrast recovery. Considering the possible artifacts & distortion for the plastic capillaries walls themselves and uneven PtG4 distribution inside, X-profiles for the period included by the two white dashed lines were averaged to obtain a more robust estimate of R. Since there was a 5° angle in between the two rods, X-profile was not accurately specified for a fixed Y-value in this case. With this definition, R = 1 represents that the two targets can be completely separated, with full contrast recovery. In the characteristics of the human vision system, it is normally assumed that two targets are distinguishable for R>0.1 (when a 10% dip in intensity is observed between them). In addition to spatial resolution, radiation dose and acquisition time were also estimated and all are listed in Table 1. Acquisition time was calculated as: "Dose(MU)" ¤"Dose rate(MU/min)", which just estimated total scanning time, excluding mechanical switching issues which vary from machine to machine. We assume that Cherenkov emission, luminescence and background signals can be read out in one pulse cycle, which is the time period between two LINAC pulses. This can be realized with a timeresolved camera by specifying the corresponding pulse delays.  In general, beam projections shaped by the MLCs are clinically defined by the targeting for the tumor region. The most unresolved shape used is a squared 'full-field' shape, with its Cherenkov image shown in Fig. 5(a). It is worth noting that, the full-field region was slightly smaller than FOV to conserve the diffused edges for both Cherenkov and luminescence images, to allow for complete measurements of the kernel function by the raw Cherenkov image for the deconvolution. For the luminescence imaging, one artifact noted was that the metal-made holder could attenuate some fluorescence emission, which led to 'slightlydimmer regions' on the PtG4 target as indicated with yellow dashed circles, but these was avoided in the analysis, and not present in the line scans analyzed, Fig. 5(d). Several conclusions can be drawn from Fig. 5. (1) Both square-and line-raster results show higher contrast recovery (R>0.15) than that of the full-field ones (R<0.1). Basically, the two capillary rods cannot be discriminated in the 'full-field' image.
(2) Image resolution seldom varies with scanning step since no space exists in between two adjacent squares, i.e., P<1. However, the recovered image would degrade a lot if P>1, e.g., the capillary rods were found to be discontinuous in Fig. 6(c) when P>1. (3) In this case, the line-scan results slightly outperformed the square point-scan ones. This could be explained by the nearly parallel direction between the capillaries placed and the illumination orientation. In other words, resolution would be reduced if the two directions were orthogonal. For example, as pointed out by the yellow arrows in Fig. 6(c), some dimmed regions could be more clearly discerned in square point-scan images, which, however, are not visible in line-scan results as shown in Fig. 6(b). These regions were highly resolved in Fig. 7, and also visible in Fig. 8. An improvement to the line-scan methods would be multi-direction scanning, which could render a full-direction resolution enhancement, however, this would come at the cost of doubling the radiation dose and acquisition time.
Multi-point scanning of squares is therefore demonstrated above to be ideal compromise between optimal resolution, minimal scan time and minimal dose. Further investigations were aimed at optimization of the pattern parameters as discussed in Section 2.A, i.e., l and P. Three kinds of square patterns were deliberately designed for this test, with their Cherenkov images displayed in greyscale as shown in Fig. 6, progressing from larger to smaller block sizes with separation of them sufficiently large to avoid any crosstalk. From the luminescence images in Fig. 5, R-value was not strongly dependent upon the value of d'. In comparison, the square size characterized by side-length l imposed a significant influence on spatial resolution, e.g., from the images (bottom line of Fig. 5) with R<0.2 when l = 77mm to those (middle line of Fig. 5) with R>0.77 when l = 14mm. However, the measurement intensity of the former cases was 5 times higher than the latter ones, which is in agreement with the conclusions of intensity versus illumination area in Section 2.1. Consequently, we applied a trade-off, where a side length of 38.3mm and P = 0.33 were tested here. In this section, the effect of increasing target depth was analyzed. It can be seen from Fig. 7 that, all the methods could provide clearly discerned images for the smallest target depth. In comparison, the full-field image presents a result of lower spatial resolution as well as SNR, and an available imaging depth of less than 8mm. For the deepest target at 13mm, only the imaging result with the square beams could provide a spatial resolution of R>0.1. Since the deblurring strategy herein only considered Cherenkov photon scattering, the recovered images for large-depth targets embedded in a scattering medium should have diffused edges. Depth-corrections for fluorescence photon scattering have been widely developed in (fluorescence) diffuse optical tomography. Besides, depth-related image deconvolution has been customized for CELSI in previous work [6].   A small target depth of 3mm was investigated to highlight the value of this deblurring by inducing a defocusing acquisition instead of photon scattering. The experimental results are shown in the Fig. 9(a), where luminescence images before and after demodulating and correction are displayed correspondingly. As references, all-in-focus images are shown in the lower line of Fig. 9. It can be seen from the defocused image before correction, the two capillaries are hard to separate (R<0.1) and even overlapped with each other at the place indicated by the yellow dashed circle. After correction, the capillaries can be discerned with R increasing to 0.26. Most EBRT treatment plans in patients dynamically shape the beam with the MLCs and Jaws within the LINAC, targeting the tumor region and minimizing dose to nearby critical structures. As a result, it was hypothesized that the natural spatial modulation done in clinical radiotherapy could be used to achieve an enhanced CELSI, based upon how the beams were formed. To investigate this, we used three typical breast treatment plans, e.g., a conformal radiation therapy (3D-CRT), an over modulated 3D-CRT (3D-CRT OM ), and an Intensity Modulated Radiation Therapy (IMRT) [30]. In addition, a dynamic wedge (DW) plan was investigated, which instead uses the Jaws slowly closing to the desirable shape. A full-size breast phanto shape transfer Alto, CA) for or 280° was u tangent for rig were 194MU (note, 100MU measurement outgoing lum positioned in to make a trad   Fig. 11(a).

Clinical geometry studies in breast scanning
The recovered images with the proposed method are shown in Fig. 11(b), with their Xand Y-profiles plotted in the left and right positions of Fig. 11(c), respectively. Significant background noise was found in the first two results, the 3D-CRT and 3D-CRT OM plans respectively, because the beamlets were less effective at excitation as shown in Fig. 11(a), leading to a reduced frame rate in the image sequence acquired. For the latter two plans, the IMRT result performed better, which was due to the combined beam shapes between a line and a square, leading to an additional improvement along Y-axis as shown in Fig. 11(c). At present, the IMRT plans are used to treat the more complex cancers, e.g., head and neck, and central nervous system [30]. For the images recovered with 3D-CRT and 3D-CRT OM plans, due to the insufficient effective inputs, targets are erroneously positioned and a bit oddly formed because of the limited beam shapes.

In-vivo mouse studies
In this section, the MDM-CELSI technique was used for 3D imaging of a tumor in a living mouse, and the pO 2 concentration was estimated for both tumor and non-tumor regions. 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. After a week of acclimatization, animals were used for injections of 10 5 MDA-MB-231 cells under the skin on one upper limb flank of the mouse in individual 0.050 ml injections. After approximately 3 weeks of growth, when a tumor with an average of 6 mm diameter was observed on the flank, the animals were then used for imaging studies. All mice were under general anesthesia of inhaled isoflurane at 1.5% in flowing air through a nose cone throughout imaging. A total of 50ml of 25μM PtG4 was directly injected into the tumor and normal regions on both blanks. For LINAC settings, a 6MV beam with a dose rate of 600MU/min was used. Dynamic MLC plans were used to generate the specified beam shapes for 3D CELSI acquisition and subsequent pO 2 estimation, respectively. In order to recover the pO2, a range of delays were created between the linac pulse and the emission collection, so that the emission intensity lifetime at each position could be recovered. Finally, this animal was imaged in the IVIS Spectrum CT system for X-ray tomography for spatial reference.

3D rendering
To render a 3D view, the beam was introduced in the lateral direction as a sheet, which was scanned bottom to top, typical of previous CELSI results, and then images captured by a vertically placed camera could be 3D stacked along depth direction [6]. A room light image under the same FOV was captured and is displayed in Fig. 12 (a). The resultant Cherenkov combined with demodulated luminescence images at each scanning positions are shown in Fig. 12(b). Similar to the scanning process in Section 4, Cherenkov beamlets swept over the target areas, which gives access to the MDM method. As a comparison, we show the raw and processed MIP images in Fig. 12(d), where the corresponding X-and Y-profiles across the tumor region is also present.   Fig. 13(b), respectively, for five time delays (t). To pursue high SNR, these measurements were measured as an integration over 100µs with a specified t after Cherenkov period. The curves of luminescence intensity against t [I(t)] for tumor and non-tumor regions were plotted in Fig. 13(c). Curve fitting procedures were performed to achieve a more substantial lifetime (τ), which was calculated according to the following equation: with I 0 being the decay initial intensity. The estimated τ values for tumor and non-tumor regions were 30.3μs and 18.1μs, with root-mean-square errors (RMSE) being 0.9 and 4.5, respectively. Then pO 2 values were calculated via the Stern-Volmer relationship: 5.6 mmHg and 76.8 mmHg for the tumor and non-tumor regions, respectively. As expected, lower pO 2 was detected for tumor region, which accords with general reports on tumor oxygen level: hypoxic or anoxic [31]. More details can be found in our previous studies on pO 2 tomography with CELSI technique [32,33]. In addition, excitation efficiency of PtG4 proportionally depends on oxygen concentration to some extent, which explains why the intensity in nontumor region is higher than that in tumor one.

Discussion and conclusion
Based on beam scan parameters examined here, and the methodology of post processing the images, a reliable enhancement approach for CELSI imaging is proposed, with an available depth of ~17mm into tissue or tissue-like medium with PtG4-dyed target. By making use of spatial control of the LINAC beam through the programmable MLCs, a modulation of the source (i.e., collimated X-ray) beam could be automatically done, as is typical for radiotherapy regimens today. Phantom experiments showed that a square block pattern generally outperformed the linescan approach, by achieving a similar spatial resolution while providing less radiation dose. The linescan has the potential for faster acquisition but the dose increase can be problematic. Further investigations about these pattern parameters with respect to image fidelity gave a recommended side length of 38mm and beam pitch of 0.33 as specified for the measurement geometry adopted in this paper. From the MTF study results, it is possible to conclude that the spatial resolution at a target depth of 18mm improves from 5.2mm to 1.8mm with the optimized beam parameters. In addition to spatial-resolution enhancement, the proposed strategy was experimentally demonstrated to be capable of out-offocus deblurring, which could be meaningful for non-contacting imaging modalities. In most fluorescence imaging modalities, the model-based excitation light fluence-rate is hard to be well predicted with experimentally or empirically estimated optical properties. For CELSI, the excitation source is from X-ray induced Cherenkov photons, which can be captured by the same device for luminescence measurement with a different pulse delay. In this sense, the proposed deconvolution method matches very well with CELSI.
For deeper applications, efforts can be made towards using more optimized oxygensensitive probes at lower concentrations [11]. At the same time, scattering-caused blurring by luminescence photons could be compensated with the image deconvolution or reconstruction strategies for diffuse optical imaging [6]. Future work will focus on applying the proposed method to improve depth-resolved CELSI with angled incidence/detection configurations.