Image derived input function for [18F]-FEPPA: application to quantify translocator protein (18 kDa) in the human brain

Abstract

In [18F]-FEPPA positron emission topography (PET) imaging, automatic blood sampling system (ABSS) is currently the gold standard to obtain the blood time activity curve (TAC) required to extract the input function (IF). Here, we compare the performance of two image-based methods of IF extraction to the ABSS gold standard method for the quantification of translocator protein (TSPO) in the human brain. The IFs were obtained from a direct delineation of the internal carotid signal (CS) and a new concept of independent component analysis (ICA). PET scans were obtained from 18 healthy volunteers. The estimated total distribution volume (V(T)) by CS-IF and ICA-IF were compared to the reference V(T) obtained by ABSS-IF in the frontal and temporal cortex, cerebellum, striatum and thalamus regions. The V(T) values estimated using ICA-IF were more reliable than CS-IF for all brain regions. Specifically, the slope regression in the frontal cortex with ICA-IF was r² = 0.91 (p<0.05), and r² = 0.71 (p<0.05) using CS-IF.

Figure 1. Internal carotid segmentation performed on OSEM-PSF images using automatic thresholding for one subject.

The red lines illustrate the lowest 36 planes containing the internal carotid artery. The black line surrounding the carotid artery represents the automatic binary mask.

Figure 2. ROIs extracted for the ICA algorithm for the same subject in Fig. 1.

Binary masks were created from the time-averaged image (first 10 frames) and applied onto each plane to automatically select the whole-brain blood regions over all planes. A) depict the arterial activity. B) depict the superior sagittal sinus.

Figure 3. Illustration of the ICA algorithm process steps.

(1) the dynamic PET images derived from sequential measurement of the radioactivity are re-arranged into 2D matrix, the first dimension refers to time acquisition fames and the second dimension refers to the spatial distribution. (2) the principal component analysis reduce matrix dimension in order to keep the most significant activity (if columns of a mixture have relatively similar TACs, then, the corresponding columns tends to be estimated as one components). (3) update the de-mixing matrix until found convergence in separation between the two components.

Figure 4. A typical double logarithmic scale of the input function estimated by CS and ICA plotted against the ABSS-IF.

Two arterial blood samples were used to correct for a small inversion at 1.5 minutes (green point) and to calibrate curves at 15 minutes post injection (blue point).

Figure 5. Bland Altman plot of total distribution volume (V_T) in the frontal cortex comparing (A) ABSS-IF versus CS-IF, and (B) ABSS-IF versus ICA-IF.

All candidates were within the 95% of limits agreed for and , with the exception of one subject for .

Figure 6. Group comparison of total distribution volume (V_T) in the frontal cortex for high affinity binders (HABs) and mixed affinity binders (MABs) calculated respectively with ABSS-IF, CS-IF and ICA-IF.

Figure 7. Illustration of the normalized histogram of the first source, ADL, GGD, and Gaussian distributions respectively.

The plot describes the data fit by three different distributions. The Gaussian distribution does not show a good model to represent data. The GGD fit better the sharper features of the histogram. However, it fails to fit well the asymmetry of data. The ADL is more appropriate to model the sharper feature of the histogram and, moreover, follows the asymmetric distribution of data through its skew parameter.