Advancing Generalisable Neural Network-Based PET Quantification: A Multicenter [¹¹C]PBR28 study

Area of Science:

• Neuroimaging
• Radiochemistry
• Artificial Intelligence

Background:

• Positron Emission Tomography (PET) quantification of tracer binding relies on accurate input function (IF) estimation.
• Arterial sampling for IF is invasive and technically challenging, limiting clinical applicability.

Purpose of the Study:

• To develop and validate a non-invasive, neural network-based framework for estimating the IF from dynamic PET data.
• To assess the generalizability of the framework across different datasets and scanners.

Main Methods:

• A patched variational autoencoder (pVAE) was used for dimensionality reduction to generate IFs with uncertainty (NNIF-dPET).
• The framework was compared against methods using image-derived input functions (NNIF-IDIF) and uncorrected blood signals (NNIF-unBlood).
• Volume of distribution (V_T) was computed using the mean output signal from the neural network-generated IFs.

Main Results:

• The NNIF-dPET method achieved accuracy comparable to traditional arterial IFs.
• NNIF-dPET outperformed methods relying on image-derived input functions.
• Latent space representations effectively approximated whole-blood activity for parent plasma IF estimation.

Conclusions:

• The developed neural network framework enables scalable and non-invasive PET quantification.
• This approach holds significant potential for broader clinical adoption of advanced PET imaging.
• Accurate IF estimation is achievable without invasive arterial sampling.