Noninvasive Arterial Input Function Estimation for Dynamic PET

35 concepts23 papers1 proposal

Positron Emission Tomography (PET) is highly valued for its ability to image molecular pathways in vivo. However, standard clinical PET scans typically yield only "semi-quantitative" snapshots, such as the Standardized Uptake Value (SUV). While convenient, SUVs conflate tracer delivery (blood flow) with actual tissue metabolism or receptor binding, and they can be heavily influenced by systemic factors like patient genetics (Yoder et al.).

To extract absolute physiological parameters—such as the metabolic rate of glucose or the precise density of neurofibrillary tangles—researchers use dynamic PET imaging. By capturing a continuous sequence of frames over time, we can track how a radiotracer moves through the body.

To turn these time-activity curves into absolute physiological units, we must solve a system of differential equations dictated by compartmental modeling. This mathematical framework treats the body as a series of interconnected pools (e.g., blood plasma, free tissue tracer, and bound tracer). Crucially, solving these equations requires knowing the system's driving force: the concentration of the tracer in the arterial blood over time. This boundary condition is known as the Arterial Input Function (AIF).

                  [ Arterial Input Function: Cp(t) ]
                                 │
                                 ▼ (K1 / k2)
                    [ Free Tissue Tracer: Cf(t) ]
                                 │
                                 ▼ (k3 / k4)
                   [ Bound/Metabolized Tracer: Cb(t) ]

The historical "gold standard" for obtaining the AIF is continuous, direct arterial blood sampling during the scan (Pascoal et al.). This requires inserting a radial artery line—an invasive, technically demanding, and risky procedure that is highly impractical for vulnerable patient populations (Wooten et al.) and logistically challenging for high-throughput or longitudinal small-animal studies (Salomonsen et al.).

To unlock the full potential of quantitative imaging, a major research direction focuses on non-invasive kinetic modeling, primarily through the automated, non-invasive estimation of the AIF.


Extracting the Input Function from the Image

The most direct alternative to drawing blood is to treat the PET scanner itself as a blood counter. If a major blood vessel lies within the scanner's field of view, we can theoretically measure the tracer concentration directly from the image voxels. This is known as an Image-Derived Input Function (IDIF).

For example, in brain imaging, researchers can draw regions of interest (ROIs) around the internal carotid arteries to track tracer delivery (Mauler et al.). In cardiac imaging, the left ventricle provides a large pool of blood from which to sample the input curve (Guérin et al.).

However, pure image extraction faces severe physical limitations:

  1. The Partial Volume Effect (PVE): PET scanners have limited spatial resolution. When a blood vessel is narrow (like the carotid artery), its signal spills over into surrounding tissues, and vice versa. This severely dampens and distorts the measured peak of the AIF.
  2. The Chicken-and-Egg Problem: To extract a clean blood signal, you need precise anatomical segmentation of the blood vessels. But to segment these structures accurately in noisy, low-resolution PET frames, you often need to know their underlying kinetic profiles (Saad et al.).

To overcome these hurdles, modern approaches combine advanced hardware with statistical unmixing. The advent of Long-Axial Field-of-View (LAFOV) PET scanners captures massive anatomical spans simultaneously at high temporal resolution, allowing AI models to locate and extract clean IDIFs across large vascular structures (Yousefirizi et al.). Mathematically, techniques like Generalized Factor Analysis of Dynamic Sequences (GFADS) can blindly separate overlapping blood and tissue signals by analyzing their distinct spatiotemporal signatures (Guérin et al.).


Cross-Modal Calibration and Reference Regions

When the PET image alone lacks the spatial resolution to yield a clean AIF, researchers look to alternative modalities or biological proxies.

Reference Region Models

For specific tracers, the need for an explicit AIF can be bypassed entirely by using a reference tissue—a region of the body known to be devoid of the target receptor or metabolic pathway. In neuroimaging, the cerebellum often serves as this reference (Wooten et al., Milella et al.). By comparing the time-activity curve of the target region to the reference region, the systemic blood concentration cancels out of the kinetic equations, allowing relative quantification without any blood sampling.

Multi-Modal Surrogates (PET-MRI)

In integrated PET-MRI systems, the structural and hemodynamic precision of MRI can be weaponized to calibrate the PET signal.

  • Contrast Dynamics: Caldeira et al. demonstrated that Dynamic Susceptibility Contrast (DSC)-MRI can track the first-pass kinetics of a magnetic contrast agent through the vasculature. Because blood flow dynamics govern both contrast agents and radiotracers, the MRI-derived curve can serve as a high-resolution surrogate to reconstruct the PET AIF.
  • Velocity-Based Scaling: Instead of tracking contrast, Fettahoglu et al. utilized phase-contrast MRI of the major neck arteries. By measuring absolute blood velocity and vessel cross-sections, the system computes a precise physical scaling factor. This factor converts relative, uncalibrated early-frame PET intensities into absolute physiological units, completely bypassing the need for invasive arterial lines.

Physics-Informed Deep Learning

The latest frontier in non-invasive kinetic modeling leverages deep learning to predict the AIF directly from dynamic image sequences (Chen et al., Salomonsen et al.). However, purely data-driven neural networks are notorious for hallucinating physically impossible curves when confronted with image degradation, noise, or motion blur.

To solve this, researchers have shifted toward physics-informed neural networks (PINNs). Instead of treating the network as a black box, the ordinary differential equations (ODEs) of tracer kinetics are embedded directly into the network's loss function.

$$\text{Loss} = \text{Loss}_{\text{data}} + \lambda \left\| \frac{dC_{\text{tissue}}}{dt} - \left( K_1 C_{\text{plasma}}(t) - k_2 C_{\text{tissue}}(t) \right) \right\|$$

This mathematical constraint acts as a regularizer. If the network predicts an AIF ($C_{\text{plasma}}$) that cannot mathematically produce the observed tissue time-activity curves ($C_{\text{tissue}}$) under the laws of biology, it is heavily penalized.

[ Dynamic PET Images ] ──► [ Deep Neural Network ] ──► [ Predicted AIF ]
                                    │                         │
                                    ▼                         ▼
                        [ Check Kinetic ODEs: dC/dt = K1*Cp - k2*C ]
                                    │
                                    ▼
                        [ Physics-Informed Penalty ]

This approach yields remarkable advantages:

  • Robustness to Artifacts: Salomonsen et al. showed that physics-informed models (PIDLIF) can accurately predict the AIF even from severely motion-blurred images, where traditional image-derived methods fail.
  • Simultaneous Parameter Mapping: By embedding these kinetic constraints into generative architectures like CycleGANs, networks can simultaneously output both the clean AIF and voxel-wise parametric maps of the entire body, eliminating the need for computationally expensive, iterative mathematical fitting (Salomonsen et al.).
  • Multi-Tracer Separation: This paradigm is even capable of "unmixing" scans where multiple radiotracers are injected simultaneously. By unrolling kinetic models into the network architecture, the physics-informed framework can isolate individual tracer kinetics and their respective input functions from a single, multiplexed scan (Pan et al., Pan et al.).

By forcing deep learning models to obey the underlying biological physics, non-invasive kinetic modeling bridges the gap between high-precision mathematical quantification and routine, patient-friendly clinical practice.

Key Foundational Work

NMF on Positron Emission Tomography (2007) — contributed 2 concept elements to this direction.
Simultaneous Segmentation, Kinetic Parameter Estimation, and Uncertainty Visualization of Dynamic PET Images (2007) — contributed 1 concept elements to this direction.
Quantitative dynamic cardiac 82Rb PET using generalized factor and compartment analyses. (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_invasive_kinetic_modeling_03295e48,
  title = {non-invasive kinetic modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-invasive-kinetic-modeling/rdr_03295e4806d344fb99b1313c22bccbf7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_invasive_kinetic_modeling_03295e48_v1,
  title = {non-invasive kinetic modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-invasive-kinetic-modeling/rdr_03295e4806d344fb99b1313c22bccbf7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake