Generation of Whole-Body FDG Parametric Ki Images from Static PET Images Using Deep Learning

FDG parametric Ki images show great advantage over static SUV images, due to the higher contrast and better accuracy in tracer uptake rate estimation. In this study, we explored the feasibility of generating synthetic Ki images from static SUV ratio (SUVR) images using three configurations of U-Nets with different sets of input and output image patches, which were the U-Nets with single input and single output (SISO), multiple inputs and single output (MISO), and single input and multiple outputs (SIMO). SUVR images were generated by averaging three 5-min dynamic SUV frames starting at 60 minutes post-injection, and then normalized by the mean SUV values in the blood pool. The corresponding ground truth Ki images were derived using Patlak graphical analysis with input functions from measurement of arterial blood samples. Even though the synthetic Ki values were not quantitatively accurate compared with ground truth, the linear regression analysis of joint histograms in the voxels of body regions showed that the mean R2 values were higher between U-Net prediction and ground truth (0.596, 0.580, 0.576 in SISO, MISO and SIMO), than that between SUVR and ground truth Ki (0.571). In terms of similarity metrics, the synthetic Ki images were closer to the ground truth Ki images (mean SSIM = 0.729, 0.704, 0.704 in SISO, MISO and MISO) than the input SUVR images (mean SSIM = 0.691). Therefore, it is feasible to use deep learning networks to estimate surrogate map of parametric Ki images from static SUVR images.

Generation of parametric Ki images for FDG PET using two 5-min scans

PURPOSE

The net uptake rate constant (K_i ) derived from dynamic imaging is considered the gold standard quantification index for FDG PET. In this study, we investigated the feasibility and assessed the clinical usefulness of generating K_i images for FDG PET using only two 5-min scans with population-based input function (PBIF).

METHODS

Using a Siemens Biograph mCT, 10 subjects with solid lung nodules underwent a single-bed dynamic FDG PET scan and 13 subjects (five healthy and eight cancer patients) underwent a whole-body dynamic FDG PET scan in continuous-bed-motion mode. For each subject, a standard K_i image was generated using the complete 0-90 min dynamic data with Patlak analysis (t* = 20 min) and individual patient's input function, while a dual-time-point K_i image was generated from two 5-min scans based on the Patlak equations at early and late scans with the PBIF. Different start times for the early (ranging from 20 to 55 min with an increment of 5 min) and late (ranging from 50 to 85 min with an increment of 5 min) scans were investigated with the interval between scans being at least 30 min (36 protocols in total). The optimal dual-time-point protocols were then identified. Regions of interest (ROI) were drawn on nodules for the lung nodule subjects, and on tumors, cerebellum, and bone marrow for the whole-body-imaging subjects. Quantification accuracy was compared using the mean value of each ROI between standard K_i (gold standard) and dual-time-point K_i , as well as between standard K_i and relative standardized uptake value (SUV) change that is currently used in clinical practice. Correlation coefficients and least squares fits were calculated for each dual-time-point protocol and for each ROI. Then, the predefined criteria for identifying a reliable dual-time-point K_i estimation for each ROI were empirically determined as: (1) the squared correlation coefficient (R² ) between standard K_i and dual-time-point K_i is larger than 0.9; (2) the absolute difference between the slope of the equality line (1.0) and that of the fitted line when plotting standard K_i versus dual-time-point K_i is smaller than 0.1; (3) the absolute value of the intercept of the fitted line when plotting standard K_i versus dual-time-point K_i normalized by the mean of the standard K_i across all subjects for each ROI is smaller than 10%. Using Williams' one-tailed t test, the correlation coefficient (R) between standard K_i and dual-time-point K_i was further compared with that between standard K_i and relative SUV change, for each dual-time-point protocol and for each ROI.

RESULTS

Reliable dual-time-point K_i images were obtained for all the subjects using our proposed method. The percentage error introduced by the PBIF on the dual-time-point K_i estimation was smaller than 1% for all 36 protocols. Using the predefined criteria, reliable dual-time-point K_i estimation could be obtained in 25 of 36 protocols for nodules and in 34 of 36 protocols for tumors. A longer time interval between scans provided a more accurate K_i estimation in general. Using the protocol of 20-25 min plus 80-85 or 85-90 min, very high correlations were obtained between standard K_i and dual-time-point K_i (R²  = 0.994, 0.980, 0.971 and 0.925 for nodule, tumor, cerebellum, and bone marrow), with all the slope values with differences ≤0.033 from 1 and all the intercept values with differences ≤0.0006 mL/min/cm³ from 0. The corresponding correlations were much lower between standard K_i and relative SUV change (R²  = 0.673, 0.684, 0.065, 0.246). Dual-time-point K_i showed a significantly higher quantification accuracy with respect to standard K_i than relative SUV change for all the 36 protocols (p < 0.05 using Williams' one-tailed t test).

CONCLUSIONS

Our proposed approach can obtain reliable K_i images and accurate K_i quantification from dual-time-point scans (5-min per scan), and provide significantly higher quantification accuracy than relative SUV change that is currently used in clinical practice.