Among the most relevant factors determining the feasibility of an IDIF approach is the vessel diameter. When large vessels are available in the FOV, IDIF extraction has been successful since the early days of PET. The thoracic aorta (inner diameter: 2.5–3 cm) and the left ventricle have been frequently employed as extraction sites for studies of the heart [17] and lung [18]. Similarly, the abdominal aorta (~ 2 cm) has been used for liver imaging [19] and the common iliac arteries (~ 1 cm) for prostate cancer studies [20]. The availability of the aorta and other large blood pools in the FOV has made IDIF the prevailing approach for input function extraction in cardiac and oncological studies, so that IDIF-based quantification of myocardial blood flow (MBF) (using 82Rb, [15O]H2O and other perfusion tracers [21]) has been successfully brought to the clinic.
IDIF extraction is instead known to be very challenging when only small-diameter, convoluted vessels, and/or with complex surrounding anatomy, are present inside the FOV. Brain studies, in particular, have had to face this issue, with the internal carotid artery having a caliber of ~ 4 to 5 mm, which is smaller than the spatial resolution of most clinical PET scanners (~ 5 mm) [3]. The limited PET spatial resolution relative to the vessel diameter introduces PVEs and spillover effects (spill-out and spill-in), which lead to loss of signal for objects smaller than 2 times the scanner’s point spread function (PSF) full width at half maximum (FWHM) [22], and can alter the amplitude and shape of the vascular signal [3]. Even in the case of the HRRT, i.e., the brain-dedicated scanner with the highest sensitivity and spatial resolution until recently, extracting a carotid IDIF free of PVE artifacts was shown not to be feasible [23]. Of course, in case of vascular pathology (e.g., arterial inflammation or trauma), the situation is made even more complex by potentially increased tracer uptake in the vessel wall, which produces spill-in effects.
To try and minimize the artifacts affecting IDIFs obtained from difficult extraction sites, multiple correction and calibration approaches have been developed, involving the use of blood samples or the estimation of recovery coefficients from knowledge of the vessel volume and scanner resolution, as detailed in [3]. The methods requiring blood samples are typically expected to be the most accurate [3]. Using venous instead of arterial samples for calibration is sometimes possible, if arteriovenous equilibrium occurs within the scan duration, with the advantage of lower invasiveness [6]. Next-generation PET scanners with LAFOV, increased sensitivity and spatial resolution, seem to be very promising for this issue [16].
The unsolved issue of radiometabolite correctionAnother main argument against IDIF approaches is that they usually still require blood sampling to account for radiometabolite activity. To obtain the proper input function (i.e., parent concentration in the arterial plasma), whole blood must first be centrifuged to separate the plasma, and the parent concentration in plasma must be separated from that of radiometabolites. With some notable exceptions, like [18F]fluorodeoxyglucose ([18F]FDG), the majority of PET tracers produce radiometabolites, which contribute to the measured radioactivity in the blood, and image-based approaches cannot distinguish them from the parent compound [3].
A population plasma parent fraction has been used, even in recent work, in case of low inter-individual variability [24]. For tracers with significant between-subject variability in the parent fraction (which can depend on various physiological or pathophysiological factors, e.g., sex, diseases, or drugs affecting hepatic function, etc. [3]), the use of a few late venous samples (when metabolite concentration is maximal) has been proposed [24, 25]. However, great caution is required because arterial and venous metabolite concentrations often differ, even at late time points [26]. Moreover, some authors have proposed using a reduced number of arterial blood samples, but of course this invalidates the non-invasiveness of the procedure [27].
Additionally, tracer concentrations in plasma and whole blood are often not the same and the plasma-over-blood (POB) ratio often varies over the duration of a PET scan [28]; in this case the use of venous samples, or a population-average curve scaled with venous samples, was shown to be an effective strategy for some tracers [24].
While the POB issue seems easier to address, the radiometabolite correction problem remains open. However, new solutions may come with the help of next-generation scanners.
The impact of motion and reconstructionMotion artifacts are known to have a significant impact on long dynamic PET acquisitions [29], and their correction is important to improve IDIF accuracy, especially when calibration with blood samples is not performed [27]. In the thorax and abdomen, motion has a stronger impact, since, unlike head motion, body motion is non-rigid. However, motion artifacts have been remarkably under-explored in the non-brain PET literature [29, 30]. This is an important issue, as body motion can result in severe misplacements of the vessel ROIs required for IDIF extraction, e.g., aorta [15].
The impact of image reconstruction on IDIF methods has been less explored. The choice of the reconstruction method (filtered back-projection, FBP vs. ordered subset expectation maximization, OSEM) may affect the quality of the IDIF curve. Scatter correction, which has been recognized as critically important to avoid biased kinetic estimates [31], is expected to impact IDIF estimation as well [3], especially at late times when the vascular tracer concentration is very low, and thus more susceptible to scatter artifacts. Additionally, noise reduction can have an important impact on IDIF recovery: The appropriate adjustment of the smoothing parameters, during and after reconstruction, is crucial to avoid biased estimates [32, 33]; moreover, advanced denoising approaches added to reconstruction have been shown to improve the match between IDIF and blood samples for [18F]FDG [34]. A critical aspect of image reconstruction is also selecting the PET frame duration, which requires a compromise between describing the rapid variations of the early IDIF and maintaining a sufficiently high signal-to-noise ratio (SNR), which has been particularly difficult for low-sensitivity scanners [35]. Next-generation scanners are expected to lead to both new opportunities and new challenges in the area of motion correction and reconstruction.
For these reasons, especially in brain PET, IDIF approaches have failed to reach widespread applicability [3]. However, developments in scanner technology in the last decade—particularly the availability of next-generation PET scanners—may be able to address these critical issues, fostering a new era for IDIF and non-invasive quantitative PET.
Automatic IDIF extraction: open-source and commercial pipelines for conventional PET scannersSignificant advances have been made during the last decade to make IDIF extraction more reproducible and reliable on conventional scanners with limited axial FOV (i.e., 15–30 cm). For instance, to obviate the variability of manual drawing of ROIs [36], automated pipelines have been developed [15, 37]. Some of these approaches have been included into fully automated routines for parametric imaging (i.e., voxel-wise mapping of the kinetic parameters of interest) (for an extensive list, see [14]).
Cardiac applications have already reached the clinic [21], thanks to the availability of large blood pools in the FOV. Specifically, multiple commercially or publicly available pipelines for quantification of MBF and MBF reserve have been developed, e.g., QPET [38], PMOD routines (http://www.pmod.com/), SyngoMBF (https://www.siemens-healthineers.com), Carimas [39], Cardiac VuEr [40], FlowQuant [41]. These pipelines are usually applicable to any cardiac perfusion tracer (i.e., 82Rb, [15O]H2O, [13N]ammonia, etc.). The software packages implement various strategies for IDIF extraction, e.g., (1) automated or semi-automated ROI placement, typically in the left ventricle, or the ascending aorta, (2) approaches based on factor analysis or clustering, or (3) hybrid approaches with ROI segmentation and factor analysis. Body and respiratory motion can potentially limit the accuracy of these methods [21, 29], as it can lead to misplacement of the ROI used for IDIF extraction, especially for longer stress studies; however, motion correction is not implemented in most of the mentioned software packages. Some studies that tested and compared the performance of these pipelines found good agreement between software packages [42, 43], but systematic evaluations of 82Rb studies on large cohorts reported significant differences in the estimates of MBF when comparing ROI-based methods with factor analysis [44]. Moreover, automated ROI placement was found to be unreliable for some of these pipelines (e.g., PMOD), with 30% failure rates, thus requiring manual adjustment [43].
Multibed–multipass imaging has made it feasible to perform whole-body dynamic PET imaging with conventional PET scanners [45], with important applications for e.g., oncology [46]. With a continuous bed motion approach, the first 5 min can be dedicated to a single-bed acquisition over the heart to capture the early IDIF kinetics, followed by multiple rapid whole-body passes to measure tissue kinetics and the IDIF tail [30]. Commercial software for whole-body parametric imaging, with a focus on [18F]FDG, has been made available by multiple vendors, including Siemens [15], GE, and United Imaging [47]. For instance, the FlowMotion MultiParametric PET suite, developed by Siemens for clinical PET/CT scanners with limited FOV (e.g., Biograph mCT, Vision 600 [48]), allows for IDIF extraction from the descending aorta or left ventricle, which are automatically identified on a low-dose computed tomography (CT) scan using the ALPHA machine learning algorithm; a ROI is placed and registered to PET images to extract the IDIF [15]. The ROI can be manually adjusted, if necessary. Although clear assessment criteria were missing, the success rate for this automated ROI placement was reported to be 95% [15], and it has already been applied in multiple studies, both with [18F]FDG [48, 49] and other tracers [50]. However, full comparison with the AIF showed that the automated ROI placement requires rigorous quality control, and performing additional motion correction is advisable [51]. It should also be noted that these whole-body acquisition routines require a compromise: if early scanning is dedicated to a chosen blood pool, full-compartmental analysis non-feasible [30].
Commercial software is not yet available for dedicated brain imaging, where the FOV does not typically include large vessels. However, a variety of open-source approaches have been developed, especially for hybrid PET/MR scanners. The first step is usually vessel segmentation, which can be performed either on anatomical images (MR, CT) or directly on early (perfusion-weighted) PET images to avoid coregistration issues [52, 53]. For PET/MRI, using time-of-flight MR angiography, in turn, can provide better vessel segmentation, thus minimizing coregistration issues [54]; moreover, concomitant MR acquisitions (for monitoring head motion) simplify PET motion correction [37, 46]; also, PET image quality may be significantly improved by using MR anatomical prior information during image reconstruction [16]. Due to a relatively longer axial FOV (e.g., 26 cm for Biograph mMR [55]), these scanners also allow imaging of larger vascular structures in the neck, including part of the common carotids (6–7 mm [56]); IDIFs extracted from cervical vessels are potentially less likely to be affected by spill-in and interindividual variability with respect to intracranial carotids [53] (Fig. 2). To further minimize PVE and spillover effects, aggressive voxel selection (e.g., via cluster analysis) [34, 53, 57] and calibration approaches [3] can be included. While IDIF calibration methods requiring one or more blood samples are expected to be the most accurate [58], PET/MRI allows for easier implementation of blood-free calibration, which relies on recovery coefficients estimated by knowing the carotid volume and scanner PSF [37]. One notable drawback of the current scenario is that IDIF validation studies in the brain have been performed mainly on healthy controls [59].
Fig. 2Comparing IDIF sites for brain PET. Comparison of three IDIF extraction sites (common carotid artery (CCA); internal carotid artery (ICA); superior sagittal sinus (SSS)) in 38 patients with glioma (brain PET acquisitions performed on a Biograph mMR scanner). IDIF curves (after the peak, fitted with a three-exponential decay model, and normalized by their maximum) shown at the individual (colored) and population mean (black) level for each extraction site (panel A: CCA, panel B: ICA, panel C: SSS). SSS had the highest between-subject variability, and CCA had the lowest. Panel D shows the mean fitted IDIFs (full-time course on the left, 20–50-min portion on the right). The curves are almost parallel, with CCA as the lowest and SSS the highest, thus suffering from highest spillover. As a note, the diameters of the three vessels are: CCA ~ 6 to 7 mm, ICA ~ 4 to 5 mm, SSS ~ 3 to 4 mm
In sum, PET/MRI IDIF methods have thus been proposed for brain PET, especially for [18F]FDG [34, 37, 53, 60] and [15O]H2O [61, 62], with satisfactory results compared to the gold-standard AIF. Multiple automatic pipelines specifically designed for PET/MR have also been made publicly available, including fully automated vessel segmentation and sophisticated blood-free partial volume correction (PVC) [37, 60]. Overall, when a rigorous pipeline for input function extraction and correction is applied, IDIF approaches have been shown to match well with gold-standard AIF in terms of parameter estimates and test–retest reliability, not only in cardiac studies [17], but also in the brain [37], exceeding the performance of PBIF and SIME [59].
IDIF application in the PET community: what is the “state of affairs”?We sought to assess how the PET community perceives IDIF approaches, both in terms of their limitations and the efforts that have been made to improve their robustness and reproducibility. To this end, a survey was disseminated to several teams with a track record of quantitative PET imaging, through a) social media (LinkedIn), b) a list of emails compiled starting from a literature search on PubMed (keywords “image-derived input function” and “PET”). The authors of the identified papers were selected after excluding publications in animal models. A copy of the survey and the full survey results are reported in Supplementary Materials. When indicated by an asterisk (*), survey participants were allowed to give more than one answer to the question.
In total, 110 researchers responded to the survey (Fig. 3, Additional file 1: Figs. S1–S4), and most (88%) had used IDIF in their studies. Survey respondents indicated they had used IDIF approaches in quantitative PET studies comparing patient populations to healthy volunteers more frequently (68%) than in methodological studies with healthy volunteers only (Fig. 3B, Additional file 1: Figs. S5, S6). This likely reflects that the applicability of IDIF has increased, perhaps due to the aforementioned automated routines for cardiac and whole-body applications. In addition, respondents used IDIF most frequently for brain PET (75%), with smaller percentages for heart, lung, liver, and whole-body PET imaging*. Notably, 18% of survey respondents also reported applying IDIF to other organs, including kidneys, intestine, prostate, breast, muscle, bone, and adipose tissue (Fig. 3C, Additional file 1: Fig. S7).
Fig. 3Survey results: participants. Information on survey respondents, with respect to years of experience in PET imaging (panel A), frequency of IDIF use in their PET studies (B), region of the body (C), and PET tracers (D) for which IDIF was applied, and overall opinion on IDIF approaches on a scale from 0 (complete disagreement) to 5 (complete agreement) (E). When not otherwise specified, the axes refer to the absolute number of answers
The most commonly used PET tracers in IDIF studies* were [18F]FDG (n = 66), [15O]H2O (n = 31), amyloid (n = 21) and TSPO (n = 19), and various oncological tracers (n = 31) (Fig. 3D, Additional file 1: Fig. S8). Thus, many respondents had used IDIF approaches when radiometabolites had no significant impact on plasma tracer concentration ([18F]FDG, [15O]H2O).
With regard to PET scanners*, respondents indicated that they performed many IDIF studies on Siemens/CTI ECAT EXACT HR + (n = 28) and Siemens Biograph mMR (n = 30) scanners, but they also reported using GE Discovery MI PET/CT (n = 19), Siemens Biograph mCT (n = 18), GE Signa PET/MR (n = 17), Siemens Biograph Vision 600 (n = 13), Siemens Biograph Vision Quadra (n = 12), Siemens HRRT (n = 12), Philips Gemini TF 64 PET/CT (n = 11), and UIH uEXPLORER total-body scanner (n = 5) (Additional file 1: Fig. S9). While these results reflect in large part the local availability of each scanner system (e.g., the older HR + was prevalent in many centers), they also highlight how newer scanners are providing the opportunity for researchers to conduct more IDIF studies; salient examples include the Biograph mMR, which led to the development of multiple IDIF pipelines for brain PET, and the next-generation scanners (Quadra, uEXPLORER).
When asked to rate IDIF approaches versus arterial sampling (Fig. 3E, Additional file
Comments (0)