Self-supervised neural network for Patlak-based parametric imaging in dynamic [¹⁸F]FDG total-body PET

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Computational Biology

Background:

• Positron Emission Tomography (PET) imaging is crucial for metabolic assessment.
• [¹⁸F]FDG PET is widely used, but full dynamic scans are time-consuming.
• Accurate kinetic parameter (K_i) imaging requires extensive data, limiting clinical utility.

Purpose of the Study:

• To develop a self-supervised neural network algorithm (SN-Patlak) for generating reliable K_i parametric images from shortened [¹⁸F]FDG total-body PET scans.
• To enable clinical applications of total-body PET by reducing scan duration.
• To validate the algorithm's performance against standard methods.

Main Methods:

• Proposed the SN-Patlak algorithm integrating neural networks with Patlak graphical analysis.
• Utilized the fitting error of the Patlak plot as the neural network's loss function.
• Employed a population-based "normalized time" to adapt the Patlak plot for shortened scans (20-50 min post-injection).

Main Results:

• SN-Patlak generated robust K_i images irrespective of dynamic PET scan duration.
• K_i images from a 10-min SN-Patlak scan (50-60 min post-injection) closely matched standard 40-min Patlak results (20-60 min post-injection).
• Quantitative comparison showed high agreement (NMSE = 0.15 ± 0.03, Pearson's r = 0.93 ± 0.01).

Conclusions:

• The SN-Patlak algorithm provides a robust and reliable method for parametric imaging.
• It enables accurate quantification from significantly shortened dynamic [¹⁸F]FDG total-body PET scans (as short as 10 minutes).
• This advancement facilitates clinical translation of total-body PET imaging.