An improved Patlak-based K_iparametric imaging approach for clinical¹⁸F-FDG total-body PET

The objective is to generate reliable Ki parametric images from 18F-FDG total-body PET with clinically acceptable scan durations using Patlak and shallow machine learning algorithms, under conditions of limited computational and data resources.

We proposed a robust and fast algorithm named Patlak-KXD to generate Ki images from dynamic PET images with shortened scan durations. In the training phase, K-means is employed to generate a Ki-balanced training dataset. Subsequently, XGBoost is utilized to learn the mapping relationship between the tissue-to-blood standardized uptake ratio (SUR) time curves and Patlak-based Ki values using this balanced dataset. In the prediction phase, the trained XGBoost can generate Ki images by calculating the Ki values from voxel-based SUR time curves obtained from the dynamic images. We compared the accuracy of Ki images generated by both the Patlak-KXD and the traditional Patlak methods across a range of shortened scan durations, and the differences in Ki images generated by the XGBoost model using static (Patlak-KXS) and dynamic PET inputs.

The Ki images generated by the Patlak-KXD from just a 4-minute (56-60 minutes) dynamic 18F-FDG total-body PET scan are comparable to those generated by the traditional Patlak method using 40-minute (20-60 minutes) dynamic PET images, as demonstrated by a normalized mean square error of 0.13 and a Pearson's correlation coefficient of 0.94 on average. The Ki images generated by the Patlak-KXD is robust to the scan duration, and the quality of Ki images generated from Patlak-KXD is superior to those from Patlak-KXS as scan duration > 10 minutes.

Reliable Ki images can be rapidly generated using shallow machine learning algorithms from dynamic 18F-FDG total-body PET scans with durations as short as four minutes. This total-body Ki parametric imaging method has potential to be used in clinical nuclear medicine and molecular imaging.