Physiological neural representation for personalised tracer kinetic parameter estimation from dynamic PET

TL;DR

The paper tackles the challenge of voxel-wise kinetic parameter estimation in dynamic $^{18}$F-FDG PET, which is computationally intensive and resolution-limited. It introduces a personalized physiological neural representation using implicit neural representations (INRs) to model dynamic PET signals and estimate the $2TC$ model parameters ($K_1$, $k_2$, $k_3$, $V_b$) per voxel, with optional integration of anatomical priors from a 3D CT foundation model. The authors demonstrate variants incorporating Gaussian Fourier Features and a SIREN network, with and without CT-derived HU inputs or foundation-model features, and evaluate on a 24-patient dynamic PET/CT dataset, comparing against a state-of-the-art DNN. Results show markedly lower voxel-wise MSE and improved spatial details in tumours and highly vascularized regions, especially for LoRes configurations, while foundation-model features offer faster convergence at higher memory cost. The work suggests that INR-based, data-efficient tracer-kinetic modelling, augmented by anatomical priors, can enhance tumour characterization, segmentation, and prognosis in clinical PET workflows.

Abstract

Dynamic positron emission tomography (PET) with [$^{18}$F]FDG enables non-invasive quantification of glucose metabolism through kinetic analysis, often modelled by the two-tissue compartment model (TCKM). However, voxel-wise kinetic parameter estimation using conventional methods is computationally intensive and limited by spatial resolution. Deep neural networks (DNNs) offer an alternative but require large training datasets and significant computational resources. To address these limitations, we propose a physiological neural representation based on implicit neural representations (INRs) for personalized kinetic parameter estimation. INRs, which learn continuous functions, allow for efficient, high-resolution parametric imaging with reduced data requirements. Our method also integrates anatomical priors from a 3D CT foundation model to enhance robustness and precision in kinetic modelling. We evaluate our approach on an [$^{18}$F]FDG dynamic PET/CT dataset and compare it to state-of-the-art DNNs. Results demonstrate superior spatial resolution, lower mean-squared error, and improved anatomical consistency, particularly in tumour and highly vascularized regions. Our findings highlight the potential of INRs for personalized, data-efficient tracer kinetic modelling, enabling applications in tumour characterization, segmentation, and prognostic assessment.

Figures (4)

• Figure 1: MSE (top row, red) and SD (lower row, blue) for a coronal slice of an exemplary patient. MMSE is higher on average over the complete slice (see FDB112 LoRes MSE Clipped vs. INR LoRes MSE), but particularly in the location of tumour (black arrow) and kidneys (green arrow), ignoring the blood pool (visible larger vessels). Also SD images show less variation on the MSE over time for our model. The comparison of LoRes and HiRes variants of our model (righ two columns) shows a good degree of agreement with more details in the HiRes image.
• Figure 2: Comparison of different variants of HiRes models in a coronal slice of an exemplary patient. The first column shows the corresponding CT and static PET images. With increased input information, the degree of granularity slightly increases while MSE average and SD range remaining almost identical.
• Figure 3: MSE along segment crossing a mediastinal lymph node metastasis (depicted in corresponding PET slice, left). The plots show the MSE as a function of space (axis going from left to bottom of the graphs) and time (axis going from bottom to right). The center plot is the MSE for the 2D baseline, while the right plot shows the inr-LoRes-2D variant. While the baseline shows the error in the center of the lesion at a late time point of the acquisition, the INR shows significantly lower MSE and rather placed at the peak of the IDIF - not visible in the graph.
• Figure 4: Coronal slice of parametric images generated with the 2D baseline (upper row second to last columns) and with the inr-LoRes-2D variant. The INR parameters show signficantly higher resolution in terms of more details and sharper edges. In concrete, the kidneys (green arrow, best visible in $K_1$ and $k_2$ images) and tumour (yellow arrow, best seen in $k_3$ image) have more clearer boundaries, capturing their heterogeneity. The tumour also shows a higher contrast to other structures pointing at higher specificity.