A Bayesian Neural Network-Based Method For The Extraction Of A Metabolite Corrected Arterial Input Function From Dynamic [11C]PBR28 PET

In PET, arterial sampling and metabolite correction are prerequisites for the gold-standard measurement of values like the volume of distribution (V
T
), often necessary for the full quantification of radioligand binding. However, the invasiveness and technical demands of these procedures limit their application in both research and clinical PET studies. Machine learning approaches have been explored to predict V
T
from PET images, but their integration in clinical routine is limited by their lack of transparency or thorough evaluation. Here we propose a Bayesian Neural Network to estimate the arterial input function (AIF), while also outputting its prediction uncertainty, 1) directly from the entire dynamic PET images (NN-AEIF), 2) from an image-derived input function (IDIF)(NN-IDIF) and, as a sensitivity measure, 3) from the un-corrected plasma curve (NN-AIF). All methods, applied on [
¹¹
C]PBR28 PET data, were compared to the metabolite-corrected AIF in terms of V
T
, and the prediction uncertainty was assessed in terms of normalised coefficient of variance (nCV). Overall, both NN-AEIF and NN-AIF were able to accurately predict V
T
, outperforming the other methods, with NN-AEIF showing the lowest nCV.