depicts an overview of the basic workflow carried out to create time-domain fluorescence image reconstructions for rodents using the FT and microCT systems. Much of the workflow is similar to the previous CW FT system, but the salient features are described briefly here. Prior to each scan, the subject (generally either a mouse or a rat) is anesthetized and placed onto the FT/CT bed and positioned onto the bite-bar in the bed’s gas anesthesia hook-up (). Motion of the subject is further restricted with tape to avoid disrupting the subject position when transiting the bed between systems. The bed is then placed into the FT system and the height of the bed is adjusted, as well as the lateral position of the subject (by moving the positions of the support rods), to center the subject in the FT imaging gantry. This is achieved by rotating the imaging gantry around the subject with the laser on, ensuring that the excitation point of the laser is always roughly normal to the circumference of the subject. Once the position of the subject on the bed has been optimized, the bed is removed from the FT system and placed in the microCT to collect a 3D anatomical image of the subject. At this stage, the CT image is used to select coronal slices of interest that can be used to guide the linear stage position in the FT system, achieved through a co-registration of the coordinate systems of the microCT and FT (the specifics of how the co-registration was carried out are discussed in Section 3.3). The bed is then moved back into the FT system and time-domain fluorescence and excitation light is collected for a customizable number of source-detector projections about each slice of interest as determined from the microCT image.
A simplified workflow depicting the full experimental procedure for carrying out time-domain fluorescence tomography in small animals.
Once collected, TPSFs (fluorescence, transmitted excitation, and reference) from each source-detector projection or time-point are background-subtracted to remove any non-time-correlated signal in the detection. Then, two calibration procedures are applied to the time-domain optical data. First, the data is corrected for differences in detector sensitivity, time delay, and pulse dispersion, as well as any filtering or differences in filter performance in each channel (Section 3.1). Second, the data is corrected for temporal changes in the laser pulse during data collection, such as drift and jitter in the intensity, timing, or dispersion. These attributes are tracked by the reference PMT (Section 3.2). The final data calibration step is to create a Born normalized data set by taking a ratio of the fluorescence to the transmittance for each source-detector projection to mitigate any inconsistencies between the data and the model used for the image reconstruction [35
]. The Born ratio data is then multiplied by a forward model of transmittance to create a normalized fluorescence data set that is inherently calibrated to the arbitrary source strength of the model [32
After FT data calibration is completed, the anatomical microCT image stack is re-incorporated into the workflow in two ways. One, it is used to localize the position of the fiberglass bed supports within the coordinate system of the FT unit in order to remove FT data from any source-detector projection that could have been attenuated by the bed (Section 3.3). Second, it is used to create a finite-element mesh of the subject, which is translated into the FT system coordinates and incorporated into a fluorescence image reconstruction algorithm as a priori anatomical information to localize the position of each source and detector relative to the subject geometry and to constrain the inverse problem (Section 3.6).
Many different image reconstruction approaches can be employed for time-domain fluorescence data, two of which are investigated in this study; however, no matter the approach, if the full benefits of the time-domain data are to be leveraged, a final pre-reconstruction step that is required is the integration of the instrument response function into the forward model used in the reconstruction algorithm. The time required for this full protocol is highly variable depending on the size of the subject, the area to be imaged, etc. A typical mouse brain scan may require about 30 minutes from start to finish: about 5 minutes for the CT scan, 15 minutes for the fluorescence/transmittance scan, with a final 10 minutes of data processing and image reconstruction.
The following subsections provide in-depth descriptions of the salient workflow procedures discussed above preceding an experimental demonstration of the utility of the workflow in phantoms in Sections 4 and 5.
3.1. System calibration and instrument response function measurement
Full calibration of a time-domain optical imaging system requires accounting for detector channel differences with respect to not only signal intensity and filter efficiency, as with continuous-wave systems, but also with respect to pulse delay and temporal dispersion of signal. The intrinsic sensitivity, pulse dispersion, and time delay properties of each detector were determined with a single calibration experiment that can be easily repeated before system use.
The calibration experiment is depicted in
and entails placing a line diffuser (Thorlabs, Newton, NJ) in the center of the imaging gantry to send equal portions of light to all detection channels. The diffuser is engineered to diffuse a collimated laser beam with angular uniformity into a fan beam spanning 100° with a width of 4°. With the detector channels spanning 90° of the imaging gantry, the excitation beam is diffused evenly into each detection channel, with negligible pulse dispersion or delay. Then, the long-pass filters in all “fluorescence” detection channels are replaced with neutral density filters. This way both the “fluorescence” channels and the “transmittance” channels are set to detect transmitted excitation light, and detector-by-detector sensitivity differences can be accounted for (see below) because every detector should have been exposed to an identical signal in terms of intensity, pulse delay, and pulse dispersion. Therefore, any observed differences can be corrected for by applying calibration factors for intensity, pulse delay, and pulse dispersion to equalize all detectors to an arbitrary detector (for the purposes of this study, the detector that measured the highest fluence was used as the reference point)—see and .
Fig. 4 System calibration and instrument response function collection. (a) A cartoon layout of the fluorescence tomography imaging system depicting the arrangement used to calibrate the system and measure the instrument response functions. Specifically, a line (more ...)
Not only can the line diffusor experiment be used to calibrate inherent intensity and time-delay differences between the different detectors, it can also be used as the instrument response function (IRFs) for each detector. By calibrating all detectors, the TPSFs from line diffusor experiment will be roughly equivalent and can be used to model laser pulse dispersion by convolving with forward model time-domain Jacobians for image reconstruction purposes. A depiction of the IRFs for all detectors is presented in (the average FWHM of the IRF is 650 ps). One caveat to this is that these IRFs are calculated for signal that has traversed the full diameter of the imaging gantry. When a specimen is placed in the middle of the gantry for imaging, the mean free path of the light outside of the specimen will be shorter than in the IRF experiment by a factor roughly equivalent to the diameter of the specimen. Therefore the calculated IRFs are shifted earlier in time by a factor of (d1 + d2)/c
, where d1
is the distance between the focal point of the source on the surface of the specimen and the center of the gantry, d2
is the distance from the middle of the gantry to the projection point of a given detector on the surface of the specimen, and c
is the speed of light in air (the typical time-shift for a mouse is approximately 100 ps). Furthermore, it should be noted that this IRF calculation is specific to the laser and PMT parameters used, so this must be kept the same between calibration and experiment. This calibration is much less time-consuming and involved than a previously described TCSPC approach [36
]. The total time required for this simplified calibration and IRF calculation is about 2 min, including setup, making it possible to run the calibration before and after every experiment for improved system characterization of the system on an experiment-to-experiment basis.
Even though this first calibration experiment provides all necessary time-domain calibrations in theory, it relies on the assumption that the engineered diffuser is aligned properly and provides an angularly uniform dispersion of the source. If Born normalization is used for image reconstruction, a more robust approach of calibrating intensity can be employed. Essentially, any type of a diffusing medium, homogeneous or not, can be placed into the center of the FT imaging gantry and both fluorescence and transmittance detection channels can be set to detect excitation light (). In this arrangement, each pair of detectors in each detection channel is expected to measure transmitted signal that is identical regardless of the shape or inhomogeneity of the diffuser used, since each pair receives light from the same collection optics. By this mean, calibration factors for each fluorescence-to-transmittance ratio measurement in each detection channel can be determined.
In general, if the same detection filter-sets are used to calibrate the system as to collect data, then it is not necessary to account for discrepancies in individual filters because they would be accounted for in the system calibration. However, since the presented calibration experiment requires both fluorescence and transmittance detection channels be set to detect excitation light, and different long-pass filter sets have to be used for detecting fluorescence using the two laser wavelength options, the inter-filter differences have to be taken into account for both calibration experiments and for data acquisition so that the calibration of signal is independent of filter differences. For the experiments presented below, the inter-filter differences were determined by placing a phantom fluorescent at both laser wavelengths (635 nm and 755 nm) was positioned at the center of the gantry and photon fluence was measured in the central detector while all filters were sequentially placed in the light path to determin inter-filter differences in efficiency.
Another key aspect to collecting robust time-domain data is being able to account for system drift over time. The laser referencing used to account for this is discussed in the following Section.
3.2. Laser reference
By monitoring the laser pulse in the FT system with a designated reference PMT, it is possible to correct for drift and jitter — in terms of intensity, pulse delay, and pulse dispersion —throughout data acquisition to improve the fidelity of the results. A major advantage to this is that the user does not have to wait for the system to warm up to a quasi-stable level prior to imaging, being able to automatically correct for variability in the system in real time. Another advantage of this setup is that by incorporating the in-line motorized attenuator (used to adjust laser power during scanning to maximize signal-to-noise characteristics without going over detection limits [33
]) upstream of the reference PMT, the laser reference signal can be used to calibrate tomography data without having to assume that the attenuation predicted by the motorized attenuator is accurate.
To demonstrate the utility of the laser reference, excitation TPSFs were collected through a line diffusor (see Section 3.1) as well as with the reference detector at 1s intervals for up to 20 minutes immediately after turning the FT system on in 2 separate experiments. In the first experiment, the laser attenuator was adjusted sequentially through 15, 10, 5, and 0 dBs at intervals of approximately 100 s to mimic an automatic exposure control experiment of an irregularly shaped specimen (
displays the intensity of one PMT detector channel and the reference detector over time). In the second experiment, TPSFs were collected in the same manner, but without adjusting the laser attenuator, and then one of the detectors was shut down and re-started at 600 s ( displays the mean-time of this detector and the reference detector over time). The results of the first experiment demonstrated that normalizing TPSFs by the signal intensity of the reference detector is a viable option for correcting for intensity drift and changes in the laser attenuator during an experiment; whereas, simply adjusting the signal by the assumed change in the laser attenuation turned out to be unreliable (). The results of the second experiment demonstrated that there could be considerable drift in the mean time of the laser pulse during warm-up: upwards of 200 ps within the first 10 min (). However, the magnitude of the shift was equivalent in all imaging detectors and in the reference detector, suggesting one, that the mean time measured by the reference detector can be used to correct for mean time drift in the imaging detectors (), and two, that, at least in this system, the major source of system drift during warm-up results from drift in the laser. Further corroborating the second point: no appreciable change in detected intensity or mean time was observed when one of the imaging detectors was shut off and turned back on at the 600 s (). After correction, the mean time drift rate in the system was −3.5 ± 20.1 ps/h. There was no appreciable change in pulse dispersion through the course of either experiment. These findings suggest that the reference PMT can be used for real-time data calibration in the same manner as the combined TPSF/IRF approach presented by Ntziachristos et al. [37
] but without having to share the total photon count of the TPSF with the IRF in each detector, thus providing a larger dynamical range.
Fig. 5 Laser referencing is illustrated here, where changes in the detected signal intensity (a) & (b) and in mean time (c) & (d) for one detection channel (blue curves) and for the laser reference channel (red curves). In (b) the normalized (more ...)
3.3. Co-registration between microCT and FT system coordinates
Specifics of the spatial coordinate registration between the microCT and the FT instruments have been described in detail previously [38
]. More concisely, a rigid-body translation is used to transform the Cartesian coordinate system of the microCT images to a polar coordinate system of the FT system to co-register the two. The choice of coordinate system is arbitrary; however, the cylindrical symmetry of the FT system makes mesh generation and source-detector placement easier in polar coordinates. The translation is repeatable and does not require fiducial markers, assuming that the local coordinate system of each instrument does not alter over time.
The translation is performed in two steps. The first is a two-dimensional translation in the plane of the FT imaging gantry and the second is a one-dimensional linear translation in the out-of-plane or stage axis. A simplified approach of co-registration was employed in this study. A 2-mm diameter fiberglass rod, fixed to the imaging bed, was positioned into the center of the imaging gantry. The positioning was verified by focusing the laser onto the rod and rotating the gantry 360° to ensure that the focal point of the excitation beam was focused onto the center of the rod at every projection angle. Once the position of the rod was verified, the bed was transferred to the microCT system and imaged. Then the center of mass of the rod in a coronal slice of the CT image was used to demarcate the location of the center of imaging gantry in the FT system within the imaging plane. Furthermore the z-axis (stage axis) location of the tip of the rod was noted in both imaging system and used to co-register the out-of-plane dimension between the two systems.
With the microCT and FT system coordinates co-registered, it is possible to use the anatomical information provided by the microCT to assist FT image reconstruction. In this study, the mesh creation package provided by NIRFAST was employed [39
]. The microCT image was used to determine the outer surface of the imaged specimen and to localize the position of the bed supports to remove FT source-projection data points that had the potential to be interrupted by the supports. FT images were then reconstructed assuming homogeneous optical properties within the specimen using Born normalization to account for data-model mismatches [40
The fiberglass rods used for supports are easy to locate in the microCT images because they have relatively high x-ray attenuation compared to biological tissue, but not too high to cause imaging artifacts. Likewise, the outline of the specimen can be located in the microCT images since biological tissue has significantly higher x-ray attenuation than air. Based on these properties, microCT images can be thresholded to create a mask for the rods and a mask for the specimen (e.g.
, top of
). These two masks are then translated into the polar coordinates of the FT system and the specimen mask is separated into a user-defined number of finite elements to create the FEM mesh [41
]. All source and detector positions employed during the FT scan are then projected to the surface of the specimen mesh along a line connecting the location of the detection or source optics on the gantry and the center of the gantry. Any source or detector projection that subtends the rod mask is removed from the data set to avoid data corruption from the bed supports.
Fig. 6 The finite-element model mesh was created as illustrated here, accounting for rod disturbance in the optical measurements. (a) A cross-sectional CT image of the optical mouse phantom shown in . The bright circles correspond to the cross-section (more ...)
presents a microCT image of a mouse phantom (Caliper Life Sciences, Hopkinton, MA)—see —with 2 cylindrical inclusions that were filled with a 100 nM solution of the AlexaFluor 647 fluorescent dye (Invitrogen, Carlsbad, CA) mixed with 1% Intralipid® in water. After the phantom was imaged in the microCT it was imaged in the time-domain FT system on top of the rods. In order to test out the mesh creation and rod removal algorithm, a fluorescence image was reconstructed using a pulse-integration approach, i.e., by summing the collected time-domain data (TPSFs) at each source-detector position over all time bins. The fluorescence image reconstruction is presented in , displaying a relatively accurate representation of the location of the two fluorescent inclusions.