INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION APPENDIX REFERENCES

DYNAMIC MULTISLICE X-ray computed tomographic (CT) imaging of the pulmonary system, when combined with a suitable radiopaque tracer agent, offers the ability to measure, in vivo and in detail, regional pulmonary perfusion throughout the lungs (13, 14, 16, 25, 26). Of particular interest in pulmonary physiology are measurements of the time red blood cells (RBCs) spend at the alveolar-capillary interface and, thus, the amount of time available for gas exchange (31). Several approaches have been used to measure the capillary transit times either directly or indirectly (6): diffusing capacity at multiple O₂ concentrations, allowing calculation of global capillary blood volume and mean transit time (MTT); in vivo microscopy, allowing measurement of RBC transit times in vessels at the lung surface (32); and magnetic resonance imaging using tracer gases (10). Electron beam CT (EBCT) allows fast (50-100 ms) acquisition of data necessary for cardiac gated imaging of passage of a radiopaque dye to obtain regional perfusion. EBCT, as well as other forms of X-ray CT scanning, has been used in animal and clinical studies of perfusion and ventilation (16, 19, 25, 27).

Given their resolution limits, whole body CT scanners are not able to resolve individual capillary beds, and regions of interest (ROIs) placed in the lung field of a dynamic image sequence most certainly contain contaminating flow signals from small arteries and veins in addition to the microvascular bed. Because we are often interested in the microvascular flow and transport characteristics, we have developed a method to extract the microvascular flow signal from regional parenchymal time-attenuation curves.

In this study, our aim was to evaluate a noninvasive CT-based method for characterizing the distribution of microvascular MTT and to use this method to evaluate microvascular MTT in animals (dogs and a pig) in the supine and prone body postures. In addition, we use a Monte Carlo simulation to estimate error in the derived microvascular perfusion parameters.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION APPENDIX REFERENCES

EBCT. Scanning protocols utilize an EBCT scanner (model C-150L, Imatron, South San Francisco, CA) (4). This scanner differs from conventional X-ray CT scanners, in that a focused electron beam is swept along tungsten targets surrounding the animal, instead of mechanical rotation of an X-ray source around the animal. This design allows for fast tomographic image acquisition, with scan apertures on the order of 50-100 ms, as opposed to ~0.6- to 1.0-s scan apertures from other high-speed commercial X-ray CT scanners. The fast scan times available with EBCT allow us to obtain stop-action images of the heart and allow us to minimize lung density changes induced by the effect of cardiac motion.

Animal preparation. All animal studies were performed within guidelines for animal care adhered to by the American Physiological Society and the National Institutes of Health. The animals were studied at the University of Pennsylvania or the University of Iowa, with approval from the respective institutional Animal Care and Use Committees.

1

1

1

1

Perfusion imaging protocol. Regional perfusion measurements were made using dynamic EBCT to follow bolus radiopaque contrast passage through the lung field. The EBCT scanner was used in the high temporal resolution (flow) mode, acquiring 6 or 8 spatial levels at 10-13 time points each. The topmost tomographic slice was centered about the carina. All animals were imaged in supine and prone postures during a brief period of apnea (10-12 s) with the lungs held at 0 cmH₂O airway pressure (functional residual capacity) during imaging gated to 80% of the R-R interval of the electrocardiogram complex. Nonionic radiopaque contrast agent (0.5-1.0 ml/kg; Omnipaque 240 or 350) was injected over 2-3 s via a powered injector connected to a 7F multiple-side-hole catheter placed in the right ventricular outflow tract. Contrast agent was injected at least one heartbeat after scan initiation to ensure at least one baseline point before contrast arrival.

Perfusion image analysis. Specific ROIs in the images were sampled in two ways for analysis: 1) specific ROIs were manually chosen to include parts of larger arteries to demonstrate the partial volume effects of including larger vessels with parenchymal regions, and 2) areas of lung parenchyma were sampled automatically by the computer using an ROI of 12.5-20.0 mm, with 7- to 8-mm slice thickness. We were careful to collect data representing only the parenchymal regions within a lung field. Samples with partial volume problems associated with regions close to the mediastinal or chest wall borders (rib cage and diaphragm) and myocardium were eliminated. Several additional criteria were established to eliminate regions containing major pulmonary blood vessels (veins and arteries) and major airways. Our laboratory previously demonstrated that, at functional residual capacity, normal lung density of a supine or prone dog ranges from ~35 to 90% air content (12). To ensure that we accepted no ROIs through which major airways or major vessels traversed, we eliminated all samples with baseline lung density >90% or <40%. Also, regions that could not be fit using the algorithms described below were discarded. After the elimination process, depending on the quality of the data set, there were 500-1,500 samples per right or left lung. This process resulted in a grid of voxels, each with a value for MTT, regional air content, and position coordinates. These regions were used to analyze the distribution of MTT within the lung using univariate regression or multiple linear regression.

(1)

(2)

View larger version (31K): [in this window] [in a new window]   Fig. 1.   Sample locations of feeding pulmonary artery (input) and parenchymal (output) regions of interest on transverse plane in supine position (A) and gamma variate-fitted input and output curves corresponding to locations (B). Size of region of interest (ROI) consists of 5 × 5 pixels, where a pixel was 0.08 mm on a side. HU, Hounsfield units.

Deconvolution method. We implemented two deconvolution methods [the fast Fourier transformation (FFT) and the damped least squares (DLS) methods] to verify, through cross validation, our result of deconvolution of the measured pulmonary artery and regional parenchymal time-attenuation curves. Before the two methods were applied, the input (feeding pulmonary artery) curve data were made to be the same length as the regional parenchymal (output) curve by addition of zeroes at the end. The sampled gamma variate functions were then deconvolved.

(3)

(4)

1

(5)

View larger version (13K): [in this window] [in a new window]   Fig. 2.   Residue functions obtained by using fast Fourier transform (FFT) and damped least squares (DLS) deconvolution methods. The 2 residue functions are almost identical. Both residue functions have a bimodal nature that consists of a first sharp, narrow peak and a second more dispersed peak. AU, arbitrary units.

View larger version (14K): [in this window] [in a new window]   Fig. 3.   Fitted original output curve and 2 output curves obtained by reconvolving the 2 residue functions in Fig. 2 with the feeding arterial curve (input). The 3 curves are essentially indistinguishable from one another, lending strength to the notion that the 2 residue functions produced by the FFT and DLS deconvolution methods are correct.

Separation of the residue function and calculation of contamination. The sharp, narrow peak in the residue function is likely due to partial volume sampling of small arteries with microvascular content. Regional residue functions of samples containing a visually apparent arterial vessel exhibit virtually only the unimodal prominent sharp, narrow peak. After the regional arterial and microvascular curves had been separated from the bimodal residue function (28), we observed a decrease in the peak amplitude of the arterial residue curve from dependent to nondependent lung regions in supine, but not prone, animals (Fig. 4). In the supine posture, this is consistent with known physiology, wherein fewer arterial regions are expected in the nondependent lung, since the lung is less dense and more fully inflated, whereas lung inflation is more uniform in the prone posture.

View larger version (76K): [in this window] [in a new window]   Fig. 4.   Reduction in the supine position and uniformity in the prone position of the magnitude of the arterial (art) component of the residue function as a function of lung height. Five samples with the same area (5 × 5 pixels) are chosen in peripheral parenchymal lung from ventral to dorsal regions. ROIs were placed to avoid visibly apparent small vessels. Amp, amplitude.

(6)

Monte Carlo simulations. Random errors were superimposed on the known, measured input (pulmonary artery) and output (regional parenchymal) time-attenuation curves obtained as described above. The resulting simulated data were fit to gamma variate functions and processed via the deconvolution algorithm described above to provide microvascular MTTs for the simulated curves. The error superimposed on the known input and output signals was obtained by assessing EBCT-derived CT images using two different noise content measurement methods (see APPENDIX).

Statistical analysis. Microvascular MTTs were evaluated with respect to height from the basal regions of the lungs in the supine and prone positions. Basal regions were chosen for imaging, because this is where the greatest vertical height of lung could be found. Each animal was scanned once in each position, producing six to eight sagittal images. Microvascular MTTs were computed in a grid of ROIs, as described above, in each slice. The vertical gradient in MTT was obtained using univariate linear regression and model-independent averaging, obtained by calculating the mean MTT for dependent and nondependent regions of the lungs. The dependent region is defined as the area 5-30% of the height from the base of the lung, and the nondependent region is defined as the area 70-95% of the height from the base of the lung.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION APPENDIX REFERENCES

To demonstrate the effect of partial-volume arteries on the first sharp peak of the residue function, we selected ROIs in which we purposely included more or less of the edge of a large pulmonary artery along with parenchyma (Fig. 5A) and ROIs that included clearly visible small arteries (Fig. 5B). Figure 5 shows the positive relation between the amount of arterial component in the ROI and the magnitude of the sharp, narrow arterial peak in the associated residue function. This can be represented by a contamination index shown in Eq. 6. In Fig. 5A, inasmuch as the ROI contains a greater portion of the feeding pulmonary artery, arterial contamination increases from 6.4 to 18.9%. Also, in Fig. 5B, inasmuch as peripheral ROIs include larger visibly apparent pulmonary vessels, the contamination index is seen to increase (from 8.5 to 18.5%).

View larger version (63K): [in this window] [in a new window]   Fig. 5.   Several ROIs containing increasing portions of pulmonary artery (A) along with lung parenchyma and several other parenchymal ROIs containing small peripheral arteries of increasing size (B). Because ROIs in A and B include a greater proportion of the pulmonary artery, the magnitude of the first sharp peak in the calculated residue functions (func) increases and the contamination index increases from 6.4 to 18.90% (A) and from 8.5 to 21.7% (B).

The mean microvascular MTTs in the supine and prone postures were 3.94 ± 1.03 and 3.40 ± 0.84 (SD) s, respectively, which were significantly different (P < 0.05 by Student's 2-tailed, paired t-test). Table 1 summarizes the supine and prone vertical gradients in MTT data using univariate linear regression for all eight animals. For all eight animals in the supine position, there was a positive linear relation between MTT and height for both lungs (P < 0.001). The slopes were significantly greater than zero: P < 0.001 by t-test and P < 0.01 using Wilcoxon signed-rank test. The slope for the linear fit for six of the eight animals in the prone position was zero or negative, with two outliers showing positive slopes. Because of the strength of these outliers, we cannot say that the slopes for the animals in the prone position are significantly different from zero. Although these two animals behaved clearly differently from all others in this series, we were unable to identify a specific reason to eliminate them from the analysis. A comparison of the difference in slopes between the supine and prone positions for each animal (Table 1) showed strong support for the hypothesis that the slopes for animals in the supine position were greater than those for animals in the prone position: P < 0.01 (paired t-test) and P < 0.02 (Wilcoxon signed-rank test). We also describe the overall difference in MTT at extremes of height (Table 2). In the supine posture, there was a ~1.8-s difference in MTT compared with a <0.5-s difference in the prone posture. This difference between prone and supine was statistically significant using the nonparametric Wilcoxon signed-rank test and the paired t-test (P < 0.005 for both tests). Figure 6 shows individual regression lines. Both methods of evaluation, with the exception of two outliers, demonstrate that, despite the reduction in the ventral-dorsal differences in microvascular MTT in the prone compared with the supine posture, the microvascular MTT remained smaller for the dorsal, basal lung region independent of body posture. Microvascular MTT slopes from the linear model and the microvascular MTT difference showed no right-to-left differences in the prone or supine postures (P > 0.7 by Student's 2-tailed, paired t-test).

View larger version (20K): [in this window] [in a new window]   Fig. 6.   Slopes of height vs. mean transit times (MTTs) using a linear regression model in supine (A) and prone (B) postures. Symbols do not represent data points; rather they are line labels. In the supine posture, 8 animals show positive slopes; in the prone posture, 5 animals display negative or zero slopes and 2 outliers show positive slopes. Slopes in the prone position are flatter than slopes in the supine position.

The distributions of the microvascular MTTs in the supine and prone positions are represented by a color-coded map in Fig. 7. The color index ranged from 0 s microvascular MTT (blue) to 8 s microvascular MTT (red). In the supine posture, red pixels are dominant in the nondependent lung region, whereas green pixels are dominant in dependent regions, indicating the strong vertical gradient from dependent to nondependent regions. However, in the prone position, the middle part of the color index bar is scattered throughout the whole lung, suggesting that there is minimal vertical gradient of microvascular MTT.

View larger version (57K): [in this window] [in a new window]   Fig. 7.   Distribution of microvascular MTTs in prone (A) and supine (B) postures. Color index is shown at right: 0 s (blue) to 8 s (red). In supine posture, vertical gradient from dependent to nondependent regions is apparent; in prone posture, vertical gradient is so small that distribution of MTTs is indistinguishable from uniform.

Figure 8 shows the distribution of air content (percent air) in the supine and prone postures. The color index ranged from 0% air (blue) to 80% air (red). In the supine posture, the nondependent region is well expanded compared with the less expanded dependent lung region. This compares with the relatively uniform expansion seen in the prone position. This correlates well with previous CT observations (11) for which an earlier dynamic volumetric scanning method was used.

View larger version (55K): [in this window] [in a new window]   Fig. 8.   Distribution of air content (%air) in prone (A) and supine (B) postures. Color index is shown at right: 0% (blue) to 80% (red). Prone position shows a relatively more uniform expansion of the lung than supine posture. Trends in this physiological parameter correspond to previously reported pulmonary physiology.

For the data obtained from five dogs, Monte Carlo simulation was performed to calculate the mean of the regional microvascular MTTs by addition of a globally or a locally calculated noise distribution (see APPENDIX) to fitted original data. Figure 9 shows a regression analysis of the original microvascular MTT against microvascular MTTs derived from data contaminated by local or global noise. The straight lines were fitted to the original microvascular MTT vs. microvascular MTT with local and global noise. The points on both graphs were contributed from the data of 5 dogs, where 10 samples (ROIs) were chosen from each dog. The slope of the original microvascular MTT and microvascular MTT with local noise is close to 1 (0.99 ± 0.01, R² = 0.99). Similarly, the slope of the original microvascular MTT and microvascular MTT with global noise is close to 1 (0.96 ± 0.09, R² = 0.99). The original mean microvascular MTT of 4.27 s vs. microvascular MTT with global noise of 4.20 ± 0.05 (SE) s (P = 0.86 by Student's 2-tailed, paired t-test) and the original microvascular MTT vs. microvascular MTT with local noise of 4.23 ± 0.01 (SE) s (P = 0.79) were not significantly different. These results demonstrate that measured imaging-derived noise does not have significant effects on the regional MTTs that are determined by application of our deconvolution-based algorithm. In addition, the means of standard errors of MTTs with global noise and local noise of 0.05 and 0.01, respectively, were small (Table 3).

View larger version (17K): [in this window] [in a new window]   Fig. 9.   Slopes and correlations of MTT of original data vs. MTT with local noise (A) and MTT of original data vs. MTT with global noise (B). Data from 5 dogs were used for Monte Carlo simulations. Both slopes are insignificantly different from 1. Original MTT and MTT with local or global noise were highly correlated. Addition of local noise leads to a better result than addition of global noise. This is expected, because standard deviation in noise distribution is larger for global than for local noise.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION APPENDIX REFERENCES

In this study, we refined our methods for calculating MTT in small lung regions using fast CT with rapid bolus injection of contrast agent. This technique allows separation of the true microvascular transport of the contrast agent from contaminating small arteries. We used a Monte Carlo technique to show a low susceptibility to noise, and we demonstrated gradients in microvascular MTTs that mostly agree with other studies.

The technique for determining regional MTTs and blood flow by following washin and washout of contrast agent has several theoretical shortcomings. 1) It is assumed that any given ROI contains only parenchyma and no feeding or draining vessels. The technique we present here allows us to detect the effects of larger arteries, but effects of small veins could still affect results. We try to minimize the effect of veins by fitting our gamma variate function only to the upper 50% of the downslope of the time-attenuation curve. 2) The regional residue function includes small precapillary arteries, capillaries, and small postcapillary veins. Thus the MTT data represent lumped effects of all three components of the microvessels above; thus, although our new method provides unique regional information in vivo, the results must be interpreted with the limitations in mind. 3) It is assumed that there is little washout of contrast agent from the veins at the time of the peak contrast concentration in the region. This assumption is very difficult to test; however, in a study of the passage of labeled RBCs through isolated lungs, Beck (3) found that a correction for venous washout had a very small effect on results. However, the assumption remains largely untested in vivo.

Validation of any method for determining MTT is difficult, because there is no accepted "gold standard" against which to compare a new method. Previous attempts to measure MTTs have included methods to obtain global average MTTs from the ratio of blood volume to total flow and relatively invasive and destructive methods that access only surface vessels. Thus our approach to validation for this study was to show that results were in reasonable agreement with previous studies and that regional gradients in MTT are consistent with known vascular physiology.

In a previous preliminary study, we showed that the deconvolved residue function has a bimodal characteristic that consists of the first sharp, narrow peak, which we believed represented the partial volume averaging of small arteries, and the second smooth, broad curve, which indicated the pure microvascular portion of the residue function (28). Supporting this notion are the data obtained through deconvolving time-attenuation curves taken from visually apparent small arteries in the lung field. The deconvolved residue functions from these samples are composed almost exclusively of a sharp, narrow peak. The data (Fig. 4) showing a reduction in the arterial peak as ROI locations were moved from dependent to nondependent lung regions can be explained as follows. 1) Because of the greater number of arterial vessels per ROI in the dependent (dorsal) lung regions of the canine pulmonary vascular architecture of the dog, there are a greater number of arterial vessels per ROI in the dependent (dorsal) lung regions. 2) Because of a vertical gradient in regional lung expansion in the supine posture, the lungs are more compressed (less expanded) in the dependent regions (Fig. 8), thus resulting in a greater numbers of partial volumed small arteries per ROI in the dependent lung. 3) Flow is lower in the nondependent regions of the supine dog. In this study, the samples that included different proportions of a visibly identifiable artery along with parenchymal regions of lung were selected. Regardless of the region of the lung, the greater the visible arterial component, the larger were the contamination indexes calculated by Eq. 6. These data support the notion that the first sharp peak of the residue function represents partial-volume small arteries, in addition to microvasculature, that are not visibly apparent in the lung parenchymal ROIs. Data are consistent with known pulmonary anatomy and physiology.

Deconvolution using the FFT method is known to be sensitive to noise (1). In the present study, we introduced and applied the DLS approach to investigate whether the first sharp, narrow peak in the residue function could be generated by noise effects on the FFT method. Clough et al. (7) used the DLS method to obtain organ transport functions through deconvolution of input and output concentration curves. We utilized the DLS algorithm, because this algorithm can reduce or remove the noise generated during deconvolution by using a regularization method introduced by Tikhonov and Arsenin (30). Fitted curves of the residue function obtained by using the deconvolution method and the DLS method closely matched each other and showed bimodal characteristics. After the resultant functions from the two methods were reconvolved with the input curves, the original output curves also matched closely the two reconvolved curves. By using the FFT deconvolution method, some of the first sharp peaks of the residue function were affected by noise, and their magnitude tended to increase. However, the second broad curve of each residue function remained intact under the influence of noise. The microvascular MTTs that are calculated using only the second broad curve of the residue function after removal of the first peak remained similar between the two methods of deconvolution. The Monte Carlo approach allows the exploration of various analytic approaches that may be difficult or laborious to implement experimentally (9, 23). We used the Monte Carlo technique to estimate the error in our deconvolution-based regional microvascular MTT measures. Noise chosen locally and globally from a series of images was superimposed on the original data to create fluctuations in the simulated density measurement. There are multiple real sources for the variability of noise. Motion of the lung results in changes in the exact tissue elements imaged in each ROI. Small changes in the amount of denser tissue elements, such as vessels, airway wall, and connective tissue, change the average ROI density according to their fractional volume. Thus the effects of lung motion increase as ROI and slice thickness decrease. Our study indicates that the mean microvascular MTT calculated from the original data was 4.28 s compared with microvascular MTT with global noise, 4.20 ± 0.05 (SE) s, and microvascular MTT with local noise, 4.23 ± 0.01 (SE) s, showing that original MTT and MTT simulated with noise were not significantly different (P > 0.8). Also, regression analysis using original MTT vs. local noise MTT and original MTT vs. global noise MTT demonstrated that the slopes were 0.99 and 0.96, respectively. When several factors, such as the number of parameters used in a gamma variate curve for fitting, the degree of noise, and FFT deconvolution, are considered, the above results demonstrate the robust nature of our algorithm (including the selected FFT deconvolution method) to obtain microvascular MTTs.

The anatomic distributions of pulmonary capillary (plasma) and RBC MTTs have been measured by other investigators (5, 15). Presson et al. (21) showed, using in vivo microscopy, that RBC MTT was 1.4 times faster than plasma MTT, because RBCs travel down the center of the vessels where velocity is greatest. Both transit times depend on the cardiac output, blood volume, and capillary recruitment. For example, the higher the cardiac output, the faster are pulmonary capillary and RBC MTTs (5). According to in vivo microscopy results of Wagner et al. (32), capillary MTTs were 12.3 ± 2.5 s in the nondependent lung, 3.1 ± 0.6 s in the middle lung, and 1.6 ± 2.5 s in the dependent lung when the dogs were placed in the lateral decubitus position. In our study, microvascular MTTs in the supine and prone postures were 3.94 ± 1.03 and 3.4 ± 0.86 (SD) s, respectively. Even though exact experimental equivalence is difficult to judge between our studies and the in vivo microscopy studies because of such factors as the different methods of measurement, possible differences in cardiac output, and different body postures, it appears that both MTTs are within similar magnitudes. Other investigators, including the in vivo microscopy group of Wagner et al. (32), showed the existence of a vertical gradient of pulmonary capillary (plasma) and RBC MTTs using different animal and body positions with the exception of the prone posture (10, 17, 18, 32). Their studies demonstrated a clear gravitationally oriented difference within the lungs, with the shortest transit times in the most dependent regions.

In our study, we applied linear and nonlinear models to determine the dependent-to-nondependent differences in MTT in the supine and prone body postures. Both models showed that the dorsal-to-ventral gradient in MTT in the supine posture is significantly reduced in the prone posture in most of our animals. Vertical gradients of capillary transit times have been investigated by others (10, 32) using imaging- and non-imaging based-methods. Wagner et al. (32) measured the capillary transit times in the upper (zone 2), middle (between zone 2 and 3), and lower (zone 3) regions of the lung of dogs and demonstrated that the transit times decreased progressively from nondependent lung to dependent lung. Therefore, they concluded that the most recruitable capillary beds and fastest transit times are in the gravitationally dependent lung. Also, human studies showed that RBC transit times are shorter in the dependent region of the lung in humans studied in the lateral decubitus position.

In summary, the FFT-based deconvolution used in this study is computationally efficient and speedy. The FFT-based algorithm we have implemented is also model free, so that the residue function is not constrained to any predefined shape or form. It is unlikely that we would have been successful had we used other deconvolution techniques that assume a shape for the residue function. Supporting the validity of the EBCT-derived regional MTT measurements is the close agreement between capillary MTTs calculated from the recovered microvascular components and direct microvascular observations reported by others, as well as the small error between original data MTTs and MTTs obtained by Monte Carlo simulation. Also the vertical gradient between dependent and nondependent regions is smaller in the prone than in the supine posture. The close agreement between the EBCT-derived data and the direct observations (in vivo microscopy) of others suggests that dynamic X-ray CT imaging in conjunction with a deconvolution-based algorithm offers a unique method to noninvasively quantitate in vivo regional microvascular perfusion, volumes, and transit times throughout the lungs. More recently, subsecond multiple detector-row spiral (helical) CT scanners have been introduced with scan times approaching the times required for quantitative pulmonary function evaluation (24). The present work will readily translate to this newly emerging, more widespread technology.