Estimating Arterial Input Function for Dynamic PET via Deep Regression

Imaging-based estimation of arterial input functions (AIFs) from dynamic positron emission tomography (PET) images offers an alternative to invasive and laborious arterial sampling. However, accurate AIF estimation is contingent on precise blood pool segmentation in these dynamic PET images, which may be compromised by limited accuracy due to the partial volume effect and noise. Here, we present a deep regression framework for direct metabolite-corrected AIF estimation from dynamic PET, obviating the need for image segmentation and mitigating potential errors in imaging-based AIF estimation. Our approach first employs a variational autoencoder in a self-supervised learning task to reconstruct the input dynamic PET. We then extract features from the bottleneck layers, which serve as compact latent representations of the 3D dynamic PET images. These latent representations are combined along the temporal axis and inputted into a deep regression network to predict the AIF. We applied the proposed method to clinical data and assessed its performance using a weighted relative error (WRE) weighted by the sensitivity of kinetic modeling. The model achieved a small WRE of 1.2±0.56%, highlighting the effectiveness of the proposed model for AIF estimation and its potential for clinical application.