A Self-Supervised Neural Network with Patlak analysis for 18F-FDG Total-body PET Parametric Imaging (2024)

Abstract

241800

Introduction: The net uptake rate constant K_i parametric images generated from dynamic ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET with Patlak plot provides high sensitivity and specificity in clinical diagnosis and therapeutic evaluation in contrast to the clinically used standardized uptake value (SUV). However, the clinical applications of K_i parametric images are mainly limited to long dynamic PET acquisition or high noise levels of K_i images generated from shortened PET scan duration. Despite some neural networks algorithms have been proposed to generate K_i parametric images from short-duration PET images, but dozens or even hundreds of long-duration dynamic PET images are needed for effective training. The objective of this study is to explore the feasibility of generating reliable K_i parametric images from shortened dynamic PET.

Methods: A self-supervised neural network algorithm with Patlak graphical analysis (SN-Patlak) was proposed to generate K_i parametric images from dynamic PET images of shortened scan duration. SN-Patlak deeply integrates neural network architecture with the regular Patlak method, employing the fitting error of the Patlak equation as the neural network's loss function. Given that the full blood time activity curve (TAC) required by the Patlak plot is unobtainable from shortened dynamic PET scans, we utilize a population-based ratios of integral-to-instantaneous blood TAC as an alternative. By integrating the modified Patlak equation into the neural network, it is only necessary to input short-duration dynamic PET images of a single subject into SN-Patlak for self-supervised iterative training, enabling the generation of corresponding high-quality K_i parametric images. This approach contrasts with other neural networks that require the collection of a large number of full-duration dynamic PET images to generate K_i parametric images as labels for supervised training. Twenty-five 60-minute ¹⁸F-FDG PET scans were conducted on a uEXPLORER total-body PET-CT scanner. The dynamic PET images (volume size: 360 × 360 ×673, voxel size: 1.667 × 1.667 × 2.886 mm³, in x, y, z direction, respectively) of 92 frames: 30×2 s, 12×5 s, 6×10 s, 4×30 s, 25×60 s, 15×120 s were reconstructed using an ordered-subset expectation maximization algorithm (4 iterations, and 20 subsets) incorporating time-of-flight and point spread function modeling. The K_i images generated from PET images using Patlak plot with t*=20 min post injection were served as the gold standard for evaluation. The precision and stability of SN-Patlak in varying scan durations and data were assessed by normalized mean square error (NMSE), Pearson's correlation coefficient (Pearson's r), and peak signal-to-noise ratio (PSNR).

Results: The K_i parametric images generated by SN-Patlak and Patlak, based on dynamic PET images from 20 to 60 minutes, show high consistency (NMSE=0.01, Pearson's r=0.99, PSNR=75.88 dB). As the scan duration was shortened, the precision of the K_i parametric images generated by SN-Patlak remained stable, contrasting with the Patlak method (Fig. 1). Specifically, even when scan durations are reduced to a clinically acceptable 10 minutes (50-60 minutes), SN-Patlak consistently yields high-quality K_i images (NMSE=0.15, Pearson's r=0.93, PSNR=62.78 dB), while Patlak produces comparatively lower-quality K_i images (NMSE=5.33, Pearson's r=0.42, PSNR=47.87 dB).