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

Quantifying the volume of distribution (VT) in Positron Emission Tomography (PET) is widely considered the gold standard for assessing tracer binding. However, this process requires an accurate estimation of the tracer's input function (IF) obtained through arterial sampling and metabolite correction-procedures that are both invasive and technically demanding. To overcome these limitations, we introduce a neural network-based framework for estimating the IF directly from [11C]PBR28 dynamic PET data, exploring generalisability across datasets and scanners. The framework employs a patched variational autoencoder (pVAE) for dimensionality reduction, generating IFs with uncertainty (NNIF-dPET), and computes VT from the mean output signal. Additionally, we evaluate two alternative methods: NNIF-IDIF, which derives IFs from image-derived input functions, and NNIF-unBlood, which uses uncorrected blood signals as input. NNIF-dPET achieves accuracy comparable to true arterial IFs while outperforming IDIF-based methods, suggesting that latent space representations can effectively approximate whole-blood activity for parent plasma input function estimation, rather than relying on pre-selected voxels.Clinical Relevance-This approach highlights the potential for scalable, non-invasive PET quantification across diverse clinical settings.