Male piglets weighing 4.1 – 11.2 kg were randomly assigned to either a test (n = 11) or control (n = 8) group. All were pre-anesthetized with an intraperitoneal injection of 40 mg/kg pentobarbital followed by intravenous administration of 5 μg/kg/hr fentanyl anesthetic and 0.2 mg/kg/hr pancurioum paralytic. Intracranial pressure, aortic pressure, end tidal respiratory CO2 concentration and body temperature were continuously monitored. Lines were inserted for injection of radiopharmaceuticals and microspheres, sampling of arterial and venous blood composition and reference organ withdrawal. Both groups received reflection of the scalp, a 1.3 cm craniotomy over the left parietal cortex and removal of the dura mater. TBI was simulated on test subjects by fluid percussion injury (FPI) at the exposed cortex. Following injury the animal was transported to the Nuclear Medicine Department at Jackson Memorial Hospital while placed on a ventilator to maintain blood gas and end tidal CO2 concentrations at normal levels. Upon arrival the standard monitoring devices were reconnected. SPECT imaging began approximately 2 hours post injury with the microsphere injection for RMT immediately after. Identical protocols were followed for control and test subjects except that test subjects received TBI.
Nuclear medicine SPECT
The SPECT camera used is a Philips-ADAC Vertex Plus Hybrid PET/SPECT (ADAC Laboratories, Milpitas, CA) with two 180° opposed rotating detectors with low energy high-resolution collimators and a 5/8" NaI scintillation crystal. The software and hardware for image reconstruction was a Philips-ADAC Pegasys (Sun OS) processing workstation for reconstruction of the SPECT image data. Approximately two hours after TBI, before injection of post-injury microspheres, brain blood perfusion SPECT imaging was performed using 99mTc-ECD. Scanning was initiated ten minutes after 5 mCi intravenous injection. Projections were collected over 24 minutes with each detector rotating 180° to make a complete 360° revolution in an elliptical orbit with the piglet placed in the supine position. Image volumes were reconstructed using the Philips-ADAC filtered backprojection algorithm with axial smoothing and a Butterworth filter with an order of five and cutoff of 0.45. Image volumes consisted of 64 × 64 × 64 voxels each measuring 4.38 mm3. The detector distance ranged from 14 cm to 20 cm during rotation resulting in a spatial resolution FWHM (full width at half-max) of approximately 12 mm [11].
Ex-vivo RMT rCBF and in-vivo SPECT CBP registration
Radioactive microspheres are small (15 μm) radiolabeled latex beads that lodge in the capillaries as they pass through the organ tissue [12]. Blood flow in any particular tissue is proportional to the radioactivity emitted from the microspheres trapped in the specimen. A 100 μCi microsphere solution (either Cerium, Scandium, Niobium or Strontium) was injected immediately after SPECT. Blood withdrawal was performed for 1.5 min at a rate of 1.9 mL/min from the arterial line by a syringe pump (Harvard Apparatus, Holliston, MA). Approximately five hours after injury the subject was euthanized with an intravenous injection of KCl (4 m Eq/kg). The brain was removed and preserved in formalin allowing for the decay of the 99mTc-ECD radiopharmaceutical used during the SPECT study.
A hierarchical dissection scheme, similar to that described by Gross et al [13], for the brain is shown in Figure 1. The right and left hemispheres were each carefully dissected into 47 regions of interest according to a standard grid (Figure 1). The cerebellum and hippocampus were not included in the analysis. Brain tissue samples weighed between 0.2 and 0.5 grams to ensure sufficient numbers of microspheres in each tissue sample. Generally, the dissection scheme can be thought of as a brain mapping template for each subject and will be referred to here as either the ROI template map, or template map. The ROI template map defined the number of serial slices and a series of samples contained within each slice. Two important features of the template map were: 1) minimal ROI size in order to detect sharp changes in rCBF and 2) ROI symmetry from the left to right hemisphere facilitating comparison of the injury site to the homologous region on the contralateral side. Dissection was performed manually, guided by the ROI template map resulting in a subject specific ROI map. Serial slices were made perpendicular to the longitudinal fissure and resulted in an average slice thickness of between 6 and 7 mm. Slices were then placed on heavy weight loose-leaf paper and further cut into symmetrical, left-right, samples following the map specification. After cutting, the outline of each sample was traced while in its natural position within the serial section. Finally, each sample was weighed and its radioactivity measured by a gamma-well counter (Packard, Minaxi Autogamma 5000).
Microsphere brain samples, similar to histological samples, are obtained by first fixing the specimen in tissue protecting alginate embedding material followed by slicing serial sections using a microtome. Successful reconstruction requires that the serial slices be realigned. Microsphere methods typically result in a relatively small number of slices whose manual realignment does not present such an onerous task. In this study, only nine slices were generated for each subject. Although the amount of realignment required was minimal, the small number of slices caused a loss of alignment information along the long axis of the brain. Registration of ROI map slices to target SPECT volume slices was achieved through visually guided manual placement in PMOD (PMOD Technologies, Ltd., Zurich, Switzerland) imaging software. The SPECT volume was translated, rotated and scaled in 3D space until a valid intermodality match was found for all slices in all dimensions. The piglet cerebrum was clearly located in the context of the head profile and reference anatomy points observable in the CBP image.
Image volume manipulation and analysis
The PMOD (PMOD Technologies, Ltd., Zurich, Switzerland) medical imaging software, developed specifically for nuclear medicine studies was used. PMOD is a general tool for quantitative nuclear medicine image analysis and processing, and modeling of function and kinetics. Requirements for image volume viewing and manipulation and registration include volume loading, navigation and transformation. Viewing (PGATE) and VOI analysis (PVOI) tools are distinct modules operating under the main PMOD application. The PGATE viewer provided volume navigation simultaneously through the coronal, sagittal and transverse planes, or views. This method for displaying the brain slice images, orthogonal view, is essential for analyzing image volume data and is standard clinical practice. Orthogonal navigation works by first choosing a 2D location in one plane (e.g. transverse) and then displaying the orthogonal planes in the remaining views (e.g. coronal and sagittal). Affine geometric transformations were required to scale, rotate and translate the image volume for the ROI map registration process. PMOD applies the scaling transformation first followed by translation. Trilinear interpolation is used to interpolate pixel values when applying spatial transformations. Each subject image volume was spatially normalized to a target volume to ensure that the voxel intensity measurements originate from homologous locations in the different brains. Implementing spatial normalization requires considerations for resampling, smoothing, and geometric transformation.
Statistical parametric mapping
VOI maps of SPECT perfusion cannot be statistically analyzed in a similar manner as the rCBF data (obtained from the RMT) due to the highly correlated nature of the measurements between adjacent SPECT VOIs. In order to analyze the injury groups for significant changes in rCBF a different strategy must be taken. As in ROI statistical mapping, a univariate test is performed at each data location, which is now an image voxel (counts) rather than an ROI region rCBF (mL/100 g/min). Simple correction methods for multiple comparisons between locations assume independence of the counts data at each voxel. As a result the correction is often too severe and statistically significant voxels are lost. The statistical parametric mapping technique as implemented by SPM2 (Wellcome Department of Imaging Neuroscience, UCL, UK) analyzes volumes for group effects at the voxel level and uses statistical theory to derive accurate estimates of data independence. Performing a statistical parametric mapping analysis requires three major stages: spatial image transformation, statistical model design, and setting parameters for making inferences on the model output.
Focal TBI lesion detection
The methods outlined above were implemented to provide a framework for detecting focal TBI by measuring the spatial extent and magnitude of rCBF and CBP changes in control and injury groups. Initially, VOI maps were reconstructed from ROI maps to validate the accuracy of the SPECT CBP studies, demonstrating that CBP changes were correlated with, but not necessarily equal to, the location and magnitude of changes in the gold standard rCBF. Statistical parametric mapping was implemented to automatically detect focal changes in CBP based on differences between the groups and not a priori ROI definitions. Statistical ROI maps were created apart from SPM to validate the results. Computational tools used to perform the analyses were based on a combination of custom Matlab (Mathworks, Natick, MA) programming and image manipulation in PMOD. Statistical parametric mapping analysis was implemented through the SPM2 toolbox for Matlab.
SPECT CBP: VOI map construction from ROI maps
ROI map slices were initially contrast enhanced, digitized and stored as jpeg image files. Digitization was performed by an HP Scanjet 3507c and HP Scanning software v1.0. Each serial section was scanned at a resolution of 100 dpi providing for both detailed resolution of ROI boundaries and a small file size. Output files were 276 × 276 pixels with 256 gray shades resulting in file sizes less than 100 KB. A total of nine jpeg files were created for each brain, one for each serial slice in the ROI map as in Figure 1. ROIs from the subject ROI maps, each with volume v and rCBF x, were reconstructed as volume-of-interest (VOI) maps with volume v and counts y (R
i
(v, x) → V
i
(v, y) i = 1.88) in 3-D space using the biomedical imaging software PMOD v2.5 (PMOD Technologies, Inc). Both the Gateway (PGATE) and VOI (PVOI) modules of PMOD were used extensively during reconstruction for affine image transformation and creation of VOIs, respectively.
Each ROI map slice was loaded (opened) as an image into PGATE as the first slice in a 276 × 276 × 64 image volume resulting in 63 blank slices plus the ROI map slice image. A flowchart (Figure 2) describes the procedure. This loading strategy allowed the slice to be translated and rotated 1) independent of the other ROI map slice images for the same brain and 2) in the same physical space as the SPECT volume. Each slice was loaded into PGATE individually and then the entire volume registration was verified. ROI map slice volumes were transformed to the same physical space as SPECT volumes by specifying the voxel dimensions as 1.02 mm × 1.02 mm × 4.38 mm and the origin coordinates. After loading, ROI slice image volumes were transformed to a common space. First, a scale factor of 0.37 was applied in the x and y dimensions to match the scaling of the SPECT volume. Next the slice was rotated 180° to match the supine positioning of the animal. Each ROI map slice image was then translated to a pre-specified z coordinate for that particular slice image. Any slices that were not originally drawn with midlines parallel to the x and y axes were corrected by rotation about its center. Following registration of each slice image volume to the common space, the volume was exported to PVOI for further processing. VOI slices were created by manually tracing individual ROI boundaries in the slice image. ROIs taken from the right hemisphere of the brain were traced first in numerical order and the process then repeated for the left hemisphere. After a VOI slice was created it was exported to a VOI definition file for later registration to the SPECT image volume.
VOI map registration to SPECT image volumes
SPECT image volumes were loaded into PGATE for preprocessing. Volume scale was set to 1.5 in the x, y and z dimensions to adjust for the fractional slice thickness resulting from the brain dissection. ROI map slices were cut with a thickness between 6 mm and 7 mm, however the SPECT transaxial slice thickness was 4.38 mm. Scaling by 1.5 effectively decreased the SPECT slice thickness to 3.285 mm allowing assignment of two SPECT transaxial slices to one ROI map serial slice. The scale factor and 2:1 assignment scheme was fixed for all subjects. Following the scale operation, coarse affine transformations were performed on the image volume to align the subject's brain into the same common space as the reconstructed ROI maps. Complete VOI maps were then imported into PVOI along with the corrected SPECT image volume and the final transformations were made to visually align the VOIs to the contour of the brain in all three dimensions (Figure 3). Averaged counts, sum total counts and VOI volume (cm3) were calculated and stored for each VOI. Counts per gram for each VOI were obtained by dividing the sum total counts by VOI volume and multiplying by the inverse of cortex density (1.05 g/mL).
SPECT CBP measurement accuracy
Both the reconstruction and registration components of the mapping methodology were tested for accuracy and precision using statistical analysis. Reconstruction tests examined the results of tracing ROI maps on paper and then in PVOI. SPECT CBP measurements are not absolute indications of blood flow, unlike microsphere rCBF, due to deterioration by three major sources of error: SPECT CBP imaging artifacts, map reconstruction, and map registration. Image artifacts include the physical effects of attenuation and scatter and smoothing caused by digital sampling (pixilation) of the field of view. Map reconstruction errors occur during excising, dissection, and tracing of the brain. When the brain is excised it tends to widen laterally and shorten vertically, although this effect is minimized by the formalin treatment. Lastly, map registration errors are introduced by the operator performing registration of the VOI map to the brain in the CBP images where suboptimal conditions, such as low uptake of the radiopharmaceutical tracer, can cause uncertainty over the correct organ landmarks in the image to align to the VOI map. Also, the VOI map must be aligned at points that are a multiple of the voxel size since the actual dissected brain slices did not have the same thickness as SPECT slices. For a large voxel size, this also may introduce error if the correct position for the VOI map is thought to be at a position in between two neighboring voxels, or planes. In examining the accuracy of the SPECT CBP data, the imaging, reconstruction and registration errors were considered collectively. The strength of the linear relationship between rCBF and CBP for each subject was tested using correlation analysis.
With the exception of extremely low or high flow values, rCBF and CBP are linearly related. Therefore, a suitable model for this relationship was the standard linear model of two variables. A predictive relationship between rCBF and CBP was established for each subject by fitting 88 paired ROI-VOI measurements to a line using linear regression analysis [12]. Both rCBF and CBP were normalized using the average value to a range between zero and one to compensate for the different units of measurement [12]. If the imaging, reconstruction and registration errors were minimal, the linear model would be a good predictor of CBP given rCBF (). This equation is used to compute predicted values of CBP, CBP', given an independent value of rCBF and slope, c
1, and intercept, c
0. Values for c
1 and c
0 are derived using the experimental paired data points at each CBP-VOI and rCBF-ROI pair according to the standard least squares normal equations. From these linear regression calculations two important indicators of the rCBF and CBP relationship are obtained. A slope, c
1, of 1 coupled with an intercept,c
0, of 0 represents perfect agreement between rCBF and CBP values. Normally, these values will be less than optimal. The correlation coefficient, r, measures the degree to which there is a linear relationship between the two variables and always falls between – 1 and 1. Linear models and correlation coefficients were calculated for each subject and the entire population using the regress and corr2 functions in Matlab.
Injury group lesion effect detection with SPM2
In the previous section the ROI Maps of rCBF values were analyzed for significant changes in rCBF due to TBI. The same analysis cannot be performed on the VOI maps derived from SPECT images because the CBP measurements are not independent and therefore no simple algorithm is available for adjusting the p value at each VOI for multiple comparisons. A statistical parametric mapping analysis, through SPM2, was employed to ultimately provide reliable estimates of independence and guidelines for inferring significant effects. Image volumes generated by the hybrid PET/SPECT camera stored in the raw data file format required conversion to the Analyze image file format for importation to SPM2. Subject 17 served as a template, or target, image to which all other subjects were spatially normalized. This particular subject was chosen because it was a control, had good contrast (high counts relative to background) and was ideally positioned in the gantry during the scanning process. Also, for optimal results the normalization algorithm requires that the selected template image have similar contrast to the image from which the parameters are to be determined. The contrast should be similar so that the normalization algorithm can reliably register the SPECT volumes based on perfusion values at each voxel.
The statistical model for comparing injured and sham groups was implemented as the predefined SPM SPECT model "Compare Populations 1 scan/subject ANCoVA". This model performs voxel-wise unpaired t-tests between the control and test groups while taking the global CBP confounding effect into account using ANCOVA. Details of the SPM2 methods are described elsewhere [14]. The confounding effect, or nuisance variable in SPM2 parlance, is the background level of perfusion (counts/pixel) in the brain. The background level of perfusion was calculated as the right hemisphere average similarly to the ROI map analysis. All normalized images were loaded into SPM2 as belonging to either the test or control group. Each calculated subject covariate was entered as a single nuisance variable. ANCOVA was selected as the grand mean scaling method and no further grand mean scaling was used. The threshold for determining brain activity was relative and set, relying on general knowledge of the images, to 0.5. No sphericity correction was applied. Using the contrast manager GUI a contrast was implemented to specify the effect of interest as any significant decrease in the test group compared to the control group. = (-1, 1). No Family-Wise Error (FWE) correction was applied in addition to the Gaussian random field correction due to the relatively small number of voxels being studied coupled with the low resolution of the SPECT CBP images. The uncorrected p-value voxel level threshold was set at 0.01 [15, 16]. The extent threshold was set to three, thus avoiding spike activations, due to the small size of the brain and to more closely match the cluster size with the resolution of the SPECT camera,. This level is comparable to 5 reported in other studies using SPECT and fMRI [16, 17].
