Introduction
Dynamic positron emission tomography (PET) can be used to measure regional myocardial blood flow (MBF) non-invasively using for example [
13N]NH
3,
82Rb or [
15O]H
2O [
1‐
3]. The introduction of hybrid PET/CT scanners enables accurate diagnosis of coronary artery disease (CAD) by combining PET perfusion studies with CT coronary angiography (CTCA) [
4‐
7]. Given the short half-life of
15O and
82Rb (122 and 76 s, respectively), repeat scans are feasible within a single scanning session, enabling stress-rest CTCA protocols with a total duration of less than 30 min. In contrast to
82Rb and [
13N]NH
3, [
15O]H
2O is freely diffusible and metabolically inert. Consequently, [
15O]H
2O is an ideal tracer for quantifying MBF, as changes in myocardial tracer activity are solely dependent on MBF and are not affected by variations in extraction fraction and/or metabolic interactions [
6].
Kinetics of [
15O]H
2O can best be described using a single-tissue compartment model with parameters for MBF and, to correct for partial volume and spillover effects, perfusable tissue fraction (PTF) [
8,
9], and left ventricular (LV) and right ventricular (RV) blood volume fractions [
10]. Using least-squares fitting techniques on segmental [
15O]H
2O data, it has been shown that resulting MBF values correlated well with MBF based on microspheres using both 2-D [
8,
11] and 3-D [
12] PET data.
In order to solve the single-tissue compartment model, time-activity curves (TACs) of arterial blood and RV, C
A(t) and C
RV(t), respectively, have to be determined. It has been shown [
10‐
13] that C
A(t) can be obtained accurately from the dynamic data themselves, eliminating the need for online blood sampling. This can be achieved by drawing volumes of interest (VOIs) in ascending aorta (AA), LV or left atrium, and RV in a blood pool image obtained using [
15O]CO and transferring these VOIs to the dynamic [
15O]H
2O data [
13]. Although the LV is often used to define the arterial input function for [
15O]H
2O, previous studies have shown that the AA is preferable for [
18F]fluorodeoxyglucose (FDG) [
14]. In addition, perfusion values based on an AA image-derived input function (IDIF) were shown to correlate well with those based on online arterial sampling [
15]. Performing an additional [
15O]CO scan, however, is cumbersome and drawing VOIs is user dependent and time consuming. In addition, there is a chance of misalignment between blood pool and water scans due to patient movement. VOIs can also be drawn on the [
15O]H
2O data themselves using a frame during the first pass of the bolus, where the blood pool is best visible. Although this eliminates errors due to patient movement between scans, it remains user dependent and time consuming. Therefore, automatic methods for extracting C
A(t) and C
RV(t) directly from a dynamic scan are preferred.
Factor analysis [
16,
17] has been used to extract blood and tissue TACs from dynamic PET images. Factor analysis separates a small number of underlying and unobservable factors that define the dynamic data. Its use in combination with dynamic [
15O]H
2O scans of the heart has been shown [
18,
19], although it was found that frequent operator intervention was required. The low signal to noise ratio of [
15O]H
2O scans, however, may affect extraction of the various factors, thereby possibly affecting quantitative accuracy of MBF. Cluster analysis [
20], k-means++ clustering [
21,
22] and factor analysis of dynamic sequences [
23] can be used as alternative segmentation algorithms. There are a number of prerequisites for a segmentation algorithm to be feasible for clinical use. Firstly, it should yield blood TACs that result in MBF values, which agree well with those obtained using manually defined blood TACs. Furthermore, segmentation reproducibility should be high, i.e. each segmentation of a single data set should yield the same flow value. Finally, it should be able to segment data reliably without or only with minimal operator intervention.
When arterial and RV TACs are available, segmental MBF can be calculated. Calculating MBF for heart segments has the obvious drawback of losing all information about the distribution of MBF within those segments. As an alternative, kinetic analyses can be performed for each voxel individually, thereby generating parametric MBF images. The gold standard for kinetic analysis, nonlinear least-squares regression (NLR), is slow and very sensitive to noise, making it unsuitable for generating parametric images. The basis function method (BFM) [
24] is a much faster and less noise-sensitive method, as it linearizes the model equation and solves it for each voxel using standard linear regression applied to a limited number of predefined possible values of MBF. Its use for [
15O]H
2O scans has been reported [
25], although the high noise level of [
15O]H
2O images on dedicated 2-D PET scanners with BGO detectors essentially ruled out calculation of MBF at the voxel level, as resulting parametric MBF images were very noisy [
26]. Improved scanning statistics of current generation 3-D-only clinical PET/CT scanners [
27] utilizing LSO or LYSO detectors and faster electronics, however, might result in parametric images of improved quality. Implementation of a correction for RV spillover [
10], which has not been reported before in combination with parametric MBF images, may further improve quantitative accuracy of MBF images, especially in the septum.
The aim of this study was to develop a method for generating quantitative parametric MBF images of diagnostic quality with minimal user intervention. To this end the use of a BFM, incorporating RV spillover, was assessed after which quantitative accuracy and reproducibility of four different segmentation algorithms for definition of blood pool TACs were compared using data acquired on a clinical 3-D PET/CT scanner.
Discussion
In the present study four different algorithms for automatically segmenting blood pool TACs were compared. In addition, a BFM for generating absolute MBF parametric images, incorporating both LV and RV spillover corrections, was validated.
Agreement of average segmental MBF and CFR, derived directly from parametric images, with segmental MBF and CFR derived using NLR, was high. This indicates that it is possible to generate quantitatively accurate parametric MBF images using the BFM implementation proposed, together with manually obtained CA(t) and CRV(t). The slope of the Deming regression was 0.977 and was significantly different from 1, indicating a small but significant underestimation of MBF in parametric images. However, an underestimation as small as 2.3% in the parametric images can be considered irrelevant for clinical practice and therefore this was considered not an issue for quantification of MBF.
Generated images (Fig.
1) were of diagnostic quality. For one patient, the predefined range of possible MBF values had to be increased due to a high stress MBF. However, stress MBF > 4.5 ml×g
−1×min
−1 can be considered to be outside the clinically relevant range of MBF, as stress MBF below 1.5–2.5 ml×g
−1×min
−1 (the actual level being dependent on age) is generally considered to be ischaemic [
7], and therefore this issue should have no influence on clinical diagnosis.
Agreement between all segmentation algorithms and manually obtained blood curves was high (ICC > 0.9 for at least one number of clusters or factors), with the highest agreement obtained for cluster analysis with six clusters (Table
1). The segmentation reproducibility of each algorithm was very good to excellent with CoVs of MBF <5% for cluster analysis, <8% for FADS and, inherently, 0% for both factor analysis and k-means++. Each algorithm overestimated MBF compared to MBF based on manually defined blood pool TACs. A possible explanation is the fact that manually obtained TACs were derived from a small volume in the AA. In contrast, the segmentation algorithms included the entire LV, AA and descending aorta (Fig.
4, online resource
1). This larger volume introduces dispersion and partial volume effects in C
A(t), resulting in higher apparent MBF compared to MBF based on a manually defined AA TAC. To verify this, C
A(t) was also obtained from manually drawn ROIs over both AA and LV, which showed a similar effect on MBF (data not shown). This effect was much smaller for CFR (Table
1), indicating that relative overestimations in stress and rest MBF were similar and cancelled out when calculating CFR.
Each algorithm had its own shortcomings. Occasionally (< 5% of cases), cluster analysis failed to separate aorta from myocardium. When looking at the 3-D images, however, it was easy to determine whether the analysis succeeded or failed (online resource
2). Consequently, due to the random starting values, a failed analysis could easily be resolved by restarting the analysis using the same number of clusters. Although this meant that the method sporadically required user intervention, this was not considered a major drawback of cluster analysis due to the ease of determining the success or failure of an analysis.
An important drawback of k-means++ is that it persistently included myocardial voxels in the arterial factor when segmenting stress scans, resulting in large overestimations of both stress MBF and CFR. This problem was independent of the number of clusters used.
FADS and factor analysis were unable to segment the RV correctly, resulting in incorrect spillover corrections. Furthermore, in contrast to cluster analysis, a low threshold had to be applied to the factor images obtained with both FADS and factor analysis, resulting in noisy images and inclusion of voxels that belonged for up to 60% to other factors. The effects of this on absolute MBF values, however, were small as seen in Table
1.
As described previously, it is essential to use the correct number of factors when using factor analysis for segmentation [
18,
19]. Similar results were found in the present study. In particular, it was not possible to find a single number of factors that could be used for segmenting all patients. As frequent user intervention, i.e. manually changing the number of factors, is required to prevent erroneous results, factor analysis was not considered feasible for clinical use. El Fakhri et al. [
31] presented a method that modifies factors and factor images after analysis by penalizing overlap in factor images. This method has not been tested in the present study, as cluster analysis did not suffer from overlap in factor images and provided good results. Furthermore, post-processing was not expected to improve feasibility of factor analysis, as incorrect segmentations were not expected to be corrected by penalizing overlap.
A limitation of the present study was that, in order to prevent memory issues during analysis, data were cropped around the heart using fixed parameters. The choice of these parameters and additional pre-processing may affect the optimal number of clusters. When different pre-processing steps are incorporated, the optimal number of clusters should be reassessed. Nevertheless, as cluster analysis was insensitive to the number of clusters chosen, this effect may be small.
Conclusion
This study demonstrates that it is possible to generate good quality parametric images of absolute MBF using [15O]H2O and a clinical PET/CT scanner. This can be achieved with minimal user intervention by using an automatic definition of blood pool TACs and on a BFM including RV spillover correction for calculation of parametric MBF images. Cluster analysis with six clusters proved to be the best segmentation algorithm for automatic definition of blood pool TACs, resulting in high correlation and agreement of MBF values with those based on manually defined blood pool VOIs. Consequently, absolute MBF images, generated from a [15O]H2O scan, are now available for clinical use.
Acknowledgements
The authors would like to thank Suzette van Balen, Judith van Es, Amina Elouahmani, Femke Jongsma, Nazerah Sais en Annemiek Stiekema for scanning patients, and Dr. Gert Luurtsema, Robert Schuit, Kevin Takkenkamp and Henri Greuter for production of [15O]H2O. This work was supported financially by Philips Healthcare.