Exploring Automatic Tissue Segmentation of AI-Generated Pharmacokinetic Parameter Maps

Ziel/Aim: Parametric maps based on pharmacokinetic models (KPMs) are very valuable and rich in information, not only for differential diagnosis but also for physiology-based segmentation. Assuming that each tissue has a unique set of kinetic parameters (KP), we hypothesize that a Neural Network (NN) can use the KPMs to segment tissues that are more distinctly visible in KPMs compared to PET/CT volumes. For example, while the vasculature would be easy to segment in PET due to its high contrast and clear boundaries, structures with high metabolic activity but heterogeneous perfusion (e.g. liver and kidneys, or tumors) might be more difficult to segment in PET due to diffuse boundaries.

Methodik/Methods: We generated KPMs (k1, k2, k3 and Vb) with the NN proposed by De Benetti et al. [1] using a dataset with 23 patients with various oncological indications undergoing LAFOV 18-FDG DynamicPET (dPET) acquisition of 65 min [2]. The four KPMs are used as channels for the input of a KPM-UNet which returns the segmentations of liver, kidneys, vasculature and aorta as outputs. Ground truth label maps were generated from the CT by the TotalSegmentator [3] and are used to supervise the training of the KPM-UNet using dice focal loss. To compare, we performed the same task using the dPET as input and we evaluated the results in terms of Dice Similarity Coefficient (DSC).

Ergebnisse/Results: The KPM-based segmentation of liver, kidneys, vasculature and aorta returned a DSC of 0.94, 0.90, 0.84 and 0.87, respectively, whereas the dPET-UNet resulted in 0.91, 0.86, 0.77 and 0.82, respectively, with an increase in DSC>0.03.

Schlussfolgerungen/Conclusions: We showed that KPM-based segmentation is feasible and returns satisfactory results. However, this is only an initial step toward a more complex unsupervised NN approach to perform simultaneous segmentation and KPM generation from dPET. We hypothesize that such an approach would be able to return the segmentation of structures which are not clearly visible in other modalities (e.g. the renal cortex, medulla and pelvis).

• Literatur/References

• 1 De Benetti, MICCAI 2023
• 2 Sari, EJNMMI 2022
• 3 Wasserthal, Radiol Artif Intell 2023

Publication History

Article published online:
12 March 2025