The experiments followed the European Ethics Committee recommendations (Decree 2010/63/CEE) and were approved by the Animal Experimental Ethical Committee of the University of Antwerp, Antwerp, Belgium (ECD 2020–59, 2020–71 and 2022–39). All scans were performed on Siemens Inveon PET/CT scanners (Siemens Medical Solutions, Inc., Knoxville, USA). For mice scanned under isoflurane anesthesia (3% for induction, 1.5% for maintenance, supplied with oxygen) 10 female C57BL/6J (group 1, 8 weeks old, body weight 18.8 \(\pm\) 0.85 g) mice (Charles River, Lyon France) underwent a 2-h dynamic [18F]SynVesT-1 scan (injected dose 11.9 \(\pm\) 2.5 MBq, injected mass 5.2 \(\pm\) 0.8 nmol/kg) following i.v. tail vein tracer injection [12]. The bolus injection of the radiotracer was administered with an automated pump (Pump 11 Elite, Harvard Apparatus, USA) in a volume of 0.2 mL with a flow of 1 mL/min. A CT scan was performed after the PET scan for anatomical reference and to perform attenuation correction.

An additional group of 5 male C57BL/6J mice (group2, 10 months old, 34.4 \(\pm\) 2.42 g; Jackson Laboratories) underwent surgery to insert an AV shunt into the femoral vein and artery to perform arterial blood sampling during the PET scan. Comparison of the plasma input function to the IDIF approach based on these data was previously published [4]. Briefly, after positioning the animal on the scanner bed under isoflurane anesthesia (3% for induction, 1.5% for maintenance), the AV shunt was connected to a peristaltic pump and to a coincidence detector (Twilite, Swisstrace) to measure whole blood activity with a 1 s resolution without blood loss due to a close loop to return blood back to the animal. Mice were administered with [18F]SynVesT-1 (13.4 \(\pm\) 4.2 MBq, 4.22 \(\pm\) 2.0 nmol/kg) and scanned for 2 h with simultaneous AV shunt blood sampling, followed by a CT scan. The bolus injection of the radiotracer was administered with an automated pump in a volume of 0.2 mL with a flow of 1 mL/min. The PET scan started 10 s before the radiotracer administration to capture the rise and peak of the input function. The activity measured with the detector was decay and background corrected, and cross-calibrated with the PET scanner each experimental day. To preserve the shape of the AV shunt TAC, noise reduction was performed using non-local means filtering. Since time of anesthesia with ketamine–xylazine is shorter (between 60 and 90 min) than the necessary time to perform the AV-sunt experiment (longer than 90 min), we did not perform the AV-sunt experiment in ketamine–xylazine anesthetized mice.

A group of 10 female C57BL/6J mice (group 3, Charles River, 8 weeks old, 18.4 \(\pm\) 0.93 g) underwent dynamic PET scans using ketamine–xylazine (ketamine 150 mg/kg, xylazine 15 mg/kg, administered 30 min before tracer injection). Mice were administered with [18F]SynVesT-1 (9.15 \(\pm\) 2.5 MBq, 6.4 \(\pm\) 0.23 nmol/kg) and scanned for 1 h, following a CT scan. The bolus injection of the radiotracer was administered with an automated pump in a volume of 0.2 mL with a flow of 1 mL/min.

Dynamic PET scans were reconstructed with in-house developed OSEM list-mode reconstruction with 16 subsets and 16 iterations, considering spatially-variant resolution modeling [13]. Attenuation correction was performed using the µ-map calculated from the CT scan. Images were reconstructed using an image grid of 128 \(\times\) 128 \(\times\) 159 voxels (0.776 \(\times\) 0.776 \(\times\) 0.796 mm) in the \(x\), \(y\), and \(z\) dimensions, respectively. Dynamic scans were reconstructed with a framing of 12 frames \(\times\) 10 s, 3 \(\times\) 20s, 3 \(\times\) 30s, 3 \(\times\) 60s, 3 \(\times\) 150s, and 9 \(\times\) 300s.

Blood extraction for whole blood activity measurement and radiometabolites analysis was performed in isoflurane and ketamine–xylazine anesthetized mice by cardiac puncture. The 20 mice from group 1 and 3 were used in this experiment, plus additional 4 mice (24 in total). All mice arrived at the laboratory facilities at the same time from the same vendor (Charles River) and had the same age (8 weeks old). The 24 mice were randomize in 2 groups of 12 mice: Twelve female mice (group 4, 21.8 \(\pm\) 1.2 g) were administered with isoflurane (3% for induction, 1.5% for maintenance) 30 min before [18F]SynVesT-1 administration (7.32 \(\pm\) 3.4 MBq) and blood was extracted using cardiac puncture at 2, 10, 30 and 60 min after tracer injection, using 3 mice per time point. The same procedure was performed in 12 female mice (group 5, 22.1 \(\pm\) 0.96 g) administered with ketamine–xylazine (ketamine 150 mg/kg, xylazine 15 mg/kg) 30 min before [18F]SynVesT-1 administration (7.25 \(\pm\) 3.2 MBq). Radiometabolites processing was performed as previously described [4, 12] to obtain the population-based plasma to whole blood ratio and the radiotracer parent fraction in plasma. For correction of all input functions, we interpolate to other time points performing a linear fit to the plasma to whole blood ratio and a sigmoid fit to the radiotracer parent fraction in plasma ratio [12]. For 10-month-old mice (group 2), we used the metabolites correction calculated for those mice in Bertoglio, et al. [4].

The heart delineated IDIF was calculated as follows [3, 4, 6] in mice from group 1, 2 and 3. From the dynamic reconstruction aligned with the CT image, the heart was identified in the CT image and the PET frame with highest heart activity was determined. A spherical volume of interest (VOI) of 3.5 mm radius was drawn on the heart region and voxels within the VOI with an activity at least 50% the maximum activity within the VOI were considered to belong to the blood pool and were used to calculate the mean whole blood time activity curve (TAC).

Due to the positivity constraint of the PET data, we use non-negative matrix factorization (NMF) as the source separation method to identify the different heart regions (mice from groups 1, 2 and 3). However, since NMF converges to a local minimum, it is important to select a robust initialization method. Here we used initialization using non-negative ICA, since this method performs the best compared to a variety of other methods [14].

Initially, a matrix \(Z\) is created by ordering in columns the individual voxel TACs considered in the spherical 3.5 mm VOI. Dimensionality reduction to the \(n\) number of components that one wants to extract is performed by principal component analysis (PCA), i.e., using the matrix \(B\) with the \(n\) eigenvectors calculated from the covariance matrix of \(Z\):

After dimensionality reduction, whitening of the data is performed using the eigenvectors matrix \(E\) and diagonal matrix \(D\) from the covariance matrix of \(x\) as described in Plumbley [15]:

The non-negative ICA algorithm [15] is used with data \(z\) to obtain the demixing matrix \(W\), and source matrix \(y\), such that \(y=Wz\). Mixing matrix \(A\) in the original dimensions is calculated by undoing the whitening and dimension reduction operations:

Finally, \(_=abs(A)\) and \(_=abs(y)\) are the initialization matrices for the NMF [16]. The resulting matrices \(G\) and \(H\) (\(Z\approx GH\)) from the NMF contain the \(n\) components TACs and corresponding weights, respectively, for every voxel considered in the heart VOI. The rows of matrix \(H\) are reordered to obtain the probability images of every component.

We assume 4 possible components can be included in the heart VOI voxels: left ventricle, right ventricle, myocardium, and contamination from the liver. Depending on the delineation of the VOI region, not all components might be included. For this reason, the NMF was calculated considering increasing number of components from 2 to 4 components. By visual inspection of the peak time and shape of the components TACs, as well as the shape of the probability images, it was determined which components were included in the VOI region. If increasing the number of components from \(n\) to \(n+1\) resulted in any 2 components TACs with the same peak time and/or spanning similar high probability voxels in the probability images, \(n\) was considered as the correct number of identifiable components. The peak time of the right ventricle occurs before the peak time of the left ventricle, while peak time of myocardium occurs after the left ventricle peak time. The liver component TAC can easily be identified by its slow increase uptake.

To further assess if the obtained components are independent, the correlation of all components pairs TACs and probability maps can be performed. If the correlation between any 2 components is above a certain threshold, it can be considered that these 2 components are not independent, and therefore the number of components has to be reduced. In our case, we found empirically that a correlation coefficient threshold of 0.7 provides good results, although this threshold might need to be adjusted depending on the shape of the different components and probability maps.

The TACs of right ventricle, left ventricle, myocardium and/or liver were calculated from the average of the voxels selected from the probability images of the corresponding component. The relative probability \(_\) of each voxel \(j\) belonging to component/region \(i\) was calculated as:

where \(_\) are the elements of matrix \(H\) calculated with NMF. For region \(i\), TACs from voxels with \(_\) > 0.9 were considered to calculate the average region \(i\) TAC.

Previously, we have validated the 2-tissue compartment model (2TCM) for mouse [18F]SynVesT-1 brain kinetics [4]. In that study, 1TCM did not result in a good fit to the brain TACs. 2TCM was applied to calculate the total volume of distribution VT in the brain regions cortex (CTX), caudate putamen (CP), thalamus (TH), hippocampus (HC), and cerebellum (CB) for mice in group 1, 2 and 3. The TACs of these regions were extracted as previously described [4, 12]. The input function used in the model fit was corrected for metabolites (either under isoflurane or ketamine–xylazine) and for plasma to whole blood ratio using population-based values calculated in the blood sampling and radiometabolites analysis experiment described above.

Sample size was defined in a previous study [12] where mean volume of distribution values with a standard deviation of 10% allowed used to detect significant relative differences of 10% with a statistical power of 0.8 between groups. No animals were excluded in the analysis. Normality of the samples was confirmed with a Wilk-Shapiro test.

The TAC extracted from the manual heart delineation (IDIF), and the left ventricle TAC calculated with NMF processing (NMF-IDIF,) for mice in groups 1, and 3, were compared with manual blood samples from mice in group 4 and 5, respectively, while IDIF and NMF-IDIF where compared with AV-shunt blood activity for mice in group 2. The difference of the standard uptake value (SUV) scaled blood samples at 2, 10, 30 and 60 min was calculated for IDIF and NMF-IDIF, for both isoflurane and ketamine–xylazine anesthetize mice in groups 4 and 5. For isoflurane AV shunt experiments (group 2), the difference between the AV shunt absolute activity, and the IDIF and NMF-IDIF activity, was calculated at the PET frames time points. Additionally, the shape similarity to the AV shunt curve was quantified by calculating the correlation between the AV shunt curve and IDIF and NMF-IDIF curves.

For the brain regional 2TCM kinetic modeling of isoflurane and ketamine–xylazine scans in groups 1 and 3, the goodness of fit was quantified either using the IDIF or the NMF-IDIF. A relative goodness of fit metric, i.e., the symmetric mean absolute percentage error (sMAPE), was used to be able to compare the fit in TACs with different levels of activity. In addition, the metric was weighted (wsMAPE) by the frame duration to consider the original, variable length, PET frames duration [17]:

$$} = \frac T_ \left| }_ - C_ } \right|}} T_ \left| }_ + C_ } \right|/2}}$$

where \(_\) is the duration of PET frame \(f\), \(}_\) is the brain region TAC activity, and \(_\) is the value of the curve fit to the 2TCM model. Finally, the wsMAPE and VT calculated from the fit with IDIF or NMF-IDIF were compared using 2-way ANOVA analysis with correction for multiple comparison (Sidaks’s test) using GraphPad Prism 9.0. We report the F value (F(degrees of freedom numerator),( degrees of freedom denominator)) and the p value. Statistical significance was set at p < 0.05. GraphPad Prism 9.5 (GraphPad Software, California, USA) was used for the analysis (Table 1).