Physically Informed Neural Network for Non-Invasive Arterial Input Function Estimation In Dynamic PET Imaging

Matteo Ferrante, Marianna Inglese, Ludovica Brusaferri, Alexander Whitehead, Marco Loggia, Nicola Toschi

Keywords: PINN, AIF, separable convolutions, quantitative PET imaging

TL;DR: A physically informed neural network for the non-invasive estimation of the arterial input function in dynamic PET imaging

Abstract: The invasive measurement of the AIF for the full quantification of dynamic PET data limits its widespread use in clinical research studies. Current methods which estimate the AIF from imaging data are prone to large errors, even when based on NNs. This work aims to estimate the AIF from dynamic PET images using physically informed deep neural networks. To this end, we employ 3D convolutions where we exploit the different channels to encode time-dependent information, and exploit depthwise separable convolutional layers to significantly reduce parameter count. We find that the incorporation of prior knowledge in the form of differentiable equations allows accurate estimation of the AIF. This allows kinetic modeling which leads to good estimates of the distribution volume. This work can pave the way for removing the large invasivity constraint that currently limits quantitative PET applications.

Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: novel methodological ideas without extensive validation

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Application: Radiology

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.