Automatic extraction of forward stroke volume using dynamic PET/CT: a dual-tracer and dual-scanner validation in patients with heart valve disease

For one patient of scanner I, injected dose was not measured and two patients showed
visually identifiable motion, and these patient had to be excluded from further analyses.
Patient characteristics of the remaining patients are shown in Table 1. Hemodynamic parameters during PET are shown in Table 2. No significant differences in blood pressures and heart rates were found between

¹⁵
O-water and
¹¹
C-acetate scans on scanner II. Blood pressures were comparable between patients scanned
on scanner I and II while heart rate was significantly higher for the patients scanned
on scanner I (p??0.01).

Table 1. Patient characteristics in mean?±?SD for continuous variables or N (%) for dichotomous
variables

Table 2. Blood pressures (in mmHg) and heart rate (min
^?1
) of all patients groups

Cluster analysis was performed automatically and successfully in all remaining patients.
Analysis time was 1 min on a standard desktop PC. Average FSV values for both tracers
and imaging modalities are shown in Table 3. A significant overestimation of FSV based on PET was found for all scanners and
tracers, with the largest overestimation for scanner II.

Table 3. Average values for FSV (in mL) derived using CMR and PET

Figure 2 shows correlation between FSV
_CMR
and FSV
_PET
using arterial blood TACs for scanner I and its corresponding Bland Altman plot. A
highly significant and high correlation was found when using arterial blood TACs (r?=?0.87, p??0.001). Similar but slightly lower correlations were found for venous blood TACs
(r?=?0.74, p??0.001) or the average FSV
_PET
(r?=?0.82, p??0.001). Bland Altman analysis revealed a systematic error (p??0.001) but no proportional error (p?=?0.737). Correlation between FSV
_CMR
and FSV
_PET
for scanner II for
¹¹
C-acetate and
¹⁵
O-water and their corresponding Bland Altman plots are shown in Figs. 3 and 4, respectively, based on the arterial blood TACs. Again, high and highly significant
correlations were found when using
¹¹
C-acetate (r?=?0.87, p?=?0.001 for arterial blood; r?=?0.86, p?=?0.001 for venous blood; and r?=?0.88, p??0.001 for the average) and similar results were found when using
¹⁵
O-water (r?=?0.85, p?=?0.002 for arterial blood; r?=?0.88, p??0.001 for venous blood; and r?=?0.88, p??0.001 for the average). Bland Altman analysis revealed both a systematic (p??0.001) and proportional error (p?=?0.004) for
¹¹
C-acetate and both a systematic (p??0.001) and proportional error (p?=?0.004) for
¹⁵
O-water.

Fig. 2. Correlation (a) between FSV
_CMR
and FSV
_PET
for scanner I based on
¹¹
C-acetate and arterial blood and its corresponding Bland Altman plot (b). Correlation coefficient, slope, and intercept of the linear fit were 0.87, 0.90,
and 27.7 mL for FSV
_PET
. No significant correlation was found in Bland Altman analyses. Continuous lines indicate the line of identity and dotted lines the linear fit in (a) and the mean difference and the 95 % confidence interval in (b)

Fig. 3. Correlation (a) between FSV
_CMR
and FSV
_PET
for scanner II based on
¹¹
C-acetate and arterial blood and its corresponding Bland Altman plot (b). Correlation coefficient, slope, and intercept of the linear fit were 0.87, 1.65,
and ?16.9 mL for FSV
_PET
, and a significant correlation in Bland Altman analyses was found. Continuous lines indicate the line of identity and dotted lines the linear fit in (a) and the mean difference and the 95 % confidence interval in (b)

Fig. 4. Correlation (a) between FSV
_CMR
and FSV
_PET
for scanner II based on
¹⁵
O-water and arterial blood and its corresponding Bland Altman plot (b). Correlation coefficient, slope, and intercept of the linear fit were 0.85, 1.69,
and ?23.0 mL for FSV
_PET
, and a significant correlation in Bland Altman analyses was found. Continuous lines indicate the line of identity and dotted lines the linear fit in (a) and the mean difference and the 95 % confidence interval in (b)

Correlation between FSV
_PET
based on
¹¹
C-acetate and
¹⁵
O-water (Fig. 5) was close to unity (r?=?0.99, p??0.001) with no systematic (p?=?0.14) or proportional (p?=?0.513) difference between measurements. Repeatability coefficient for these measures
was 11.0 mL.

Fig. 5. Correlation (a) between FSV
_PET
based on
¹¹
C-acetate and based on
¹⁵
O-water when using arterial blood TACs and its corresponding Bland Altman plot (b). Correlation coefficient, slope, and intercept of the linear fit were 0.99, 1.03,
and ?6.4 mL, respectively, and no correlation was found in the Bland Altman plot.
Repeatability coefficient was 11.0 mL. Continuous lines indicate the line of identity and dotted lines the linear fit in (a) and the mean difference and the 95 % confidence interval in (b). RPC repeatability coefficient

Discussion

This study shows the feasibility of a fully automated method of measuring forward
stroke volume using the indicator-dilution principle and dynamic PET with two different
tracers and scanners. The method requires no additional manual labor or separate PET
reconstructions over those required for standard quantitative analysis of dynamic
PET data. When using the arterial blood time-activity curve, a high correlation (r???0.85) was found with FSV as measured with the gold standard, phase-contrast CMR.

The method uses cluster analysis 18], 19] for extraction of image-derived input functions (i.e. C_A
(t)). This approach minimizes interobserver variability in quantitative analysis of
myocardial blood flow 4], 18] and allows for integration of FSV
_PET
measurements in a clinical workflow without additional workload and independent of
operator skill level. The method presented in this study is routinely applicable to
any dynamic cardiac PET study, provided that a standardized and rapid infusion protocol
is used and scan data with at least 90 s of short time frames is available.

There were no significant differences between the values obtained with
¹⁵
O-water and
¹¹
C-acetate (p?=?0.14). This illustrates the consistency of the method between different tracers
and its high reproducibility. In addition, factors such as the increased positron
range and shorter half-life of
¹⁵
O or differences in uptake patterns between
¹⁵
O-water and
¹¹
C-acetate did not affect the obtained results. This suggests that the method can be
used with any tracer, as long as it is injected as a rapid bolus in a standardized
fashion. In this study, automated injection devices were used in both centers and
the infusion protocol was identical, eliminating possible biases due to differences
in injection methodology. To reproduce the current results, care has to be taken to
avoid a bolus injection that is too rapid as the low time-sampling typically found
in PET studies limits the accuracy of boluses with too steep time-activity curves.
Bolus infusion times of less than 5 s are therefore not recommended. In addition,
care has to be taken that the infusion time is fast enough to minimize overlap between
the first and second pass of the bolus through the blood pool. Including the second
(or higher) pass in C_A
(t) will lead to overestimations of the area under the curve of the first-pass peak
and consequently an underestimation of FSV
_PET
.

On the other hand, a systematic bias with CMR was found for both scanners (p??0.001), and this bias was different between scanners. Differences in for instance
scatter corrections, detector material and crystal dimensions, or counting performance
during the first-pass might induce scanner-dependent differences in obtained FSV values.
In addition, the influence on partial volume effects is expected to be scanner dependent,
manifesting itself in differences in contrast recovery. Furthermore, it has been shown
that using reconstructions with corrections for the point-spread function (such as
the TrueX reconstruction used for group I) yield higher contrast recovery as compared
to standard iterative reconstructions 20]. Increased contrast recovery leads to increased areas under the peak of the first-pass
and consequently less overestimation of FSV values. This might explain some of the
differences in results between the scanners used in this study. Standardized corrections
for partial volume effects using scanner-dependent recovery coefficients might reduce
or even eliminate some of these issues. However, aortic diameter is not consistent
between cardiac patients and using a fixed (a priori) recovery coefficient might introduce
different biases in the data as these coefficients are object-size dependent. Since
correlation with FSV based on CMR was high for both scanners, scanner-specific correction
factors can be derived and used to get consistent values independent of scanner used.
Alternatively, when this is not possible, FSV
_PET
can be used in a relative fashion when patients are scanned multiple times, as reproducibility
of the method is high (Fig. 5).

Irrespective of these scanner differences, both scanners yielded a systematic overestimation
of FSV. Underestimations of the total blood activity due to the partial volume effect
(PVE), which is especially prominent in the descending aorta and to a lesser extent
in the ascending aorta, lead to overestimations of FSV. To evaluate the impact of
PVE, one additional erosion step was applied to the obtained clusters, keeping only
the most central voxels which should suffer less from partial volume effects. For
scanner I, no differences in FSV were found as compared to values obtained without
additional erosion (p?=?0.13), suggesting that partial volume effects were not significant. On the other
hand, for scanner II, FSV values were significantly lower after an additional erosion
(4.0?±?1.6 %, p??0.001), showing that partial volume effects still play a minor but significant
role for this older scanner. Nevertheless, other factors such as scanner counting
performance at high dead times, scatter corrections, and accuracy of the injected
dose remain that might play a role and could be investigated further. However, the
high correlation coefficients observed in this study show the consistency of the bias,
and potentially correction factors can be applied for a more routine setting on current
generation of PET/CT scanners. This however requires further validation.

Phase-contrast CMR is a versatile and thoroughly validated technique, at least theoretically
not associated with scanner-related or operator bias although this has not been documented.
The protocols might differ, as in this study, which potentially accounts for some
differences in mean FSV
_CMR
between the two cohorts. In this study, there were three main differences in CMR protocol
between sites. First, for patients investigated with scanner I, flow velocity was
measured at the level of the LVOT because in patients with a stenotic aortic valve,
flow is not laminar downstream of the aortic valve, and therefore flow measurements
would inevitably be subjected to larger variation if the sampling site was chosen
to be the ascending aorta. In patients investigated with scanner II, flow was measured
in the ascending aorta. However, these patients did not have a stenotic aortic valve,
flow was assumed to be laminar in the ascending aorta and consequently, the impact
of this difference in CMR protocol is expected to be marginal. Second, patients investigated
with scanner I had CMR during breathhold, while patients in the second cohort had
CMR during free breathing. Cardiac loading conditions change with the breathing cycle,
and, as a consequence, the measured stroke volume differs. All PET scans were done
with free breathing, and the bolus travels through either chamber during a few breaths.
The excellent repeatability of the PET measurements performed with scanner II suggests
that free breathing during first-pass PET does not induce significant errors. However,
if PET is calibrated to CMR for the sake of measurement portability, the actual CMR
protocol might affect the PET values. Finally, the temporal resolution was higher
in patients scanned on scanner II (32 cardiac phases per beat) as compared to scanner
I (15 cardiac phases per beat). For this reason, it is expected that FSV
_CMR
estimates are more precise and accurate for the former group although systematic differences
are not expected. Finally, it has to be noted that differences in age, body-surface
area, heart rate, loading conditions, and disease state between both groups may account
for some differences in average FSV.

FSV measurements of the patients scanned on scanner II had a more pronounced bias,
compared to CMR. These patients had significant mitral regurgitation, potentially
enhancing the bias. Phase-contrast CMR separates outward and backward flow velocities
while the PET method only measures net outflow of the bolus. The PET application of
FSV measurements is an implementation of standard indicator dilution technique, similar
to the invasive thermodilution approach, and also successfully applied for decades
using radionuclides and external gamma counters 21]. PET has a much lower time resolution than both temperature probes and gamma cameras,
but since the clusters (Fig. 1) incorporate both the left-ventricular and the left-atrial cavity, regurgitated blood
is still included in the arterial cluster. This will increase the area under the first-pass
curve proportionally, and mitral regurgitation is therefore accounted for when estimating
FSV
_PET
. This was confirmed by the excellent correlation between FSV
_CMR
and FSV
_PET,
as well as the excellent correlations between FSV
_PET
for both venous and arterial blood. Consequently, the bias observed with scanner II
is related to the PET device and not the cardiac condition or the indicator dilution
principle per se.

The method applied in this study utilizes cluster analysis; an automatic segmentation
method which is use in cardiac PET has been shown before 18] and is successfully applied in a clinical setting 4]. This is in contrast to other studies using manually defined blood TACs 14], 22], 23] or semi-automated methods 24], 25]. The automated segmentation method has the obvious advantage of reducing or eliminating
observer-induced variation and increasing workflow. Since arterial input functions
are fundamental for any quantitative PET study, the method described in this study
requires no extra work other than entering the heart rate during the start of the
scan and the injected dose. The proposed method is therefore easily applied to a clinical
routine. In addition, isolation of the first-pass peak from C_A
(t) or C_V
(t) is performed fully automatically. In contrast to 23], no assumptions are made regarding the upslope of the peak, and except for the exponential
fit starting after the frame with the maximum downslope (t₂
), the original data of C_A
(t) and C_V
(t) are retained. Fewer processing steps to the data are likely to limit potential biases
and errors due to post-processing.

There are several limitations to the method described in this study. First, the injected
dose must be administered as a rapid bolus in order to accurately isolate the peak
of the first-pass. In cases where the bolus is fractionated or where a significant
part of the bolus is for instance stuck in the arms, the relationship between the
injected dose and the area of the first-pass peak (Eq. 1) is compromised and results
are unreliable. Similarly, when the injection is performed as a slow bolus or an infusion,
as is typically the case when using
⁸²
Sr/
⁸²
Rb generators, overlap between the first and second pass of the tracer can occur and
will influence obtained values of FSV. Consistent tracer administration methods are
therefore required for accurate measurements of FSV. The patients included in this
study all had valvular abnormalities. Because of this, a comparison between regular
stroke volume based on CMR or ECG-gated PET data and FSV
_PET
was not performed, as significant differences between total stroke volume and forward
stroke volume cannot be ruled out.

Left-ventricular ejection fraction (LVEF) may be a more powerful predictor of cardiac
events than SV alone. However, LVEF requires not only SV but also end-diastolic volume
which cannot obtained using the methods described in this study, ruling out calculation
of LVEF. More studies are warranted for automated and routine extraction of end-diastolic
volume from dynamic PET images, enabling subsequent calculation of LVEF.

This study shows the validity of FSV
_PET
for
¹¹
C-acetate and
¹⁵
O-water. In spite of the highly different downstream distribution pattern, these tracers
performed equally well for measuring FSV. This suggests that the new method is directly
applicable to most other and more widely used tracers, such as
¹⁸
F-FDG and
¹³
N-ammonia.
⁸²
Rb-rubidium generators may have variable infusion schemes based on generator age,
which will require additional validation of FSV measurements in that setting. Similarly,
additional validation is also needed for tracers with potentially very high first-pass
lung retention, such as
¹¹
C-PIB in studies of cardiac amyloidosis and some tracers of highly lipophilic drugs.
However, for some of these tracers, absolute quantification is not routinely performed,
and static uptake images are used instead, often accompanied with ECG-gated reconstructions.
Performing separate dynamic reconstructions to obtain FSV
_PET
has little value over the use of regular SV based on ECG-gated reconstructions in
these cases. The use of FSV
_PET
can therefore only be recommended for tracers which frequently require dynamic scanning
and absolute quantification, as the method can be incorporated in the standard analysis
only in these cases.