A robust and versatile deep learning model for prediction of the arterial input function in dynamic small animal $\left[^{18}\text{F}\right]$FDG PET imaging

TL;DR

This paper introduces FC-DLIF, a fully convolutional deep-learning model that predicts the arterial input function ($AIF$) directly from dynamic small-animal PET data, obviating invasive blood sampling. It employs a two-stage architecture with a 3D ResNet-based spatial feature extractor and a 1D temporal feature extractor, enabling input-length flexibility and robustness to time shifts. Quantitative results show FC-DLIF outperforms a baseline DLIF in $AIF$ estimation and improves downstream kinetic parameters, though generalization to unseen tracers remains a challenge and may benefit from transfer-learning strategies. The approach promises to streamline preclinical PET workflows, reduce animal use, and support longitudinal studies by providing a non-invasive, efficient $AIF$ predictor with potential for real-time deployment.

Abstract

Dynamic positron emission tomography (PET) and kinetic modeling are pivotal in advancing tracer development research in small animal studies. Accurate kinetic modeling requires precise input function estimation, traditionally achieved via arterial blood sampling. However, arterial cannulation in small animals like mice, involves intricate, time-consuming, and terminal procedures, precluding longitudinal studies. This work proposes a non-invasive, fully convolutional deep learning-based approach (FC-DLIF) to predict input functions directly from PET imaging, potentially eliminating the need for blood sampling in dynamic small-animal PET. The proposed FC-DLIF model includes a spatial feature extractor acting on the volumetric time frames of the PET sequence, extracting spatial features. These are subsequently further processed in a temporal feature extractor that predicts the arterial input function. The proposed approach is trained and evaluated using images and arterial blood curves from [$^{18}$F]FDG data using cross validation. Further, the model applicability is evaluated on imaging data and arterial blood curves collected using two additional radiotracers ([$^{18}$F]FDOPA, and [$^{68}$Ga]PSMA). The model was further evaluated on data truncated and shifted in time, to simulate shorter, and shifted, PET scans. The proposed FC-DLIF model reliably predicts the arterial input function with respect to mean squared error and correlation. Furthermore, the FC-DLIF model is able to predict the arterial input function even from truncated and shifted samples. The model fails to predict the AIF from samples collected using different radiotracers, as these are not represented in the training data. Our deep learning-based input function offers a non-invasive and reliable alternative to arterial blood sampling, proving robust and flexible to temporal shifts and different scan durations.

Figures (14)

• Figure 1: The proposed architecture consists of two parts: a 3D ResNet acts as a spatial feature extractor (SFE), and a 1D convolutional network acts as a temporal feature extractor (TFE). The first takes in input the data volume of each time step and reduces its dimensions to one vector of 32.0 features. Then the extracted vectors of all time steps are stacked along the time dimension, over which the second part of the model performs a series of convolutions, until the vector dimension is reduced to 1. The final output of the model is a time-series with the same length as the input data.
• Figure 2: Quantitative comparison between the baseline and the proposed model over each sample, (a) boxplot of the MSE from the two models; (b) Lineplot with mean and 95% confidence interval for different Poisson noise factors.
• Figure 3: Scatterplot summarizing the results over the whole dataset for the arterial input function estimation for each model. The points show the pointwise comparison between true and predicted values for each mouse, represented by 42.0 time steps for both the baseline and proposed models. The axes labels indicate the SUV for either the predicted (SUV$_\mathrm{DLIF}$), or measured input function (SUV$_\mathrm{AIF}$). Ideally for a perfect fit, all points would lie on the black dashed line $y=x$, with $a=1$ and $b=0$. The coefficient of determination $r^2\leq1$ quantifies how well the predictor fits the data. Similarly, Pearson's correlation coefficient $r\leq 1$ measures the linear correlation between the AIF and the DLIF.
• Figure 4: Scatterplot summarising the results over the whole dataset for voxel-wise tracer kinetic modeling. Each point represents a voxel within a mouse, randomly sampled over 50000.0 voxels. Also for this figure, the fitted lines should ideally be overlapping the dashed $y=x$ line
• Figure 5: Examples of input function predicted with FC-DLIF and compared against the ground truth AIF. The insets zoom on the first 3.0 minutes of the curves, to emphasize the tracer uptake peak. (a) Best sample; (b) median sample; (c) worst sample.