FDG PET Scan Durations via Effective Data Processing.

BACKGROUND
Total-body dynamic PET (dPET) imaging using 18 F-fluorodeoxyglucose (18 F-FDG) has received widespread attention in clinical oncology. However, the conventionally required scan duration of approximately 1 hour seriously limits the application and promotion of this imaging technique. In this study, we investigated the possibility and feasibility of shortening the total-body dynamic scan duration to 30 min post-injection (PI) with the help of a novel Patlak data processing algorithm for accurate Ki estimations of tumor lesions.


METHODS
Total-body dPET images acquired by uEXPLORER (United Imaging Healthcare Inc.) using 18 F-FDG of 15 patients with different tumor types were analyzed in this study. Dynamic images were reconstructed into 25 frames with a specific temporal dividing protocol for the scan data acquired 1 hour PI. Patlak analysis-based Ki parametric imaging was conducted based on the imaging data corresponding to the first 30 min PI, during which a Patlak data processing method based on cubit Hermite interpolation (THI) was applied. The resultant Ki images acquired by 30-min dynamic PET data and the standard 1-hour Ki images were compared in terms of visual imaging effect, region signal-to-noise ratio (SNR), and Ki estimation accuracy to evaluate the performance of the proposed Ki imaging method with a shortened scan duration.


RESULTS
With the help of Patlak data processing, acceptable Ki parametric images were obtained from dynamic PET data acquired with a scan duration of 30 min PI. Compared with Ki images obtained from unprocessed Patlak data, the resulting images from the proposed method performed better in terms of noise reduction. Moreover, Bland-Altman (BA) plot and Person correlation coefficient (PPC) analysis showed that that 30-min Ki images obtained from the processed Patlak data had higher accuracy for tumor lesions.


CONCLUSION
Satisfactory Ki parametric images with high tumor accuracy can be acquired from dynamic imaging data corresponding to the first 30 min PI. Patlak data processing can help achieve higher Ki imaging quality and higher accuracy regarding tumor lesion Ki values. Clinically, it is possible to shorten the dynamic scan duration of 18 F-FDG PET to 30 min to acquire an accurate tumor Ki and further effective tumor detection with uEXPLORER scanners. This article is protected by copyright. All rights reserved.


Introduction
Positron emission tomography (PET) is an important medical tool for oncological diagnosis and therapy guidance [1][2][3][4]. Two classic PET image acquisition frameworks are usually employed [5]: (a) static PET imaging and associated standardized uptake value (SUV), which are mainly used in clinical practice; or (b) in contrast to this conventional setting, dynamic PET imaging that yields multi-frame image series to enable the recovery of more detailed metabolic information, which plays a key role in achieving more precise clinical diagnoses, tumor staging and treatment plan formulations [6][7][8][9][10]. With the most commonly used PET tracer, 18Ffluorodeoxyglucose ( 18 F-FDG), dynamic PET scanning generally lasts for one hour after tracer injection since images need to be acquired after the tracer is fully cleared from normal tissues. However, such a long scan duration has severely limited the clinical demand for dynamic imaging. The cost of imaging is high because of the long exposures required for the detectors and the large amount of data collected that requires processing. Furthermore, this leads to significantly reduced patient comfort level, which may lead to patient motion that results in image artifacts.
Therefore, shortening the required total scan duration of 18 F-FDG dynamic PET imaging is of great importance. Some investigations on shortening scan durations have been reported in recent years. Most works in this field are based on three shortened dynamic scanning protocols: (i) only late-time dynamic scanning: these studies lack data corresponding to the early stage; thus, an alternate plasma input function (IF) is needed, such as population-based IF (PBIF) [11][12][13][14]; (ii) dual-scan imaging, which requires two separate dynamic scans with short durations after tracer injection that act as complements to each other [15][16][17][18]; and (iii) only early-time scanning, with dynamic analysis is carried out based on the acquired data from the early frames [19][20][21]. Using the first protocol, the scan duration is effectively reduced, and a static image search for clinical diagnosis can be obtained. However, the alternate IF may introduce error into the dynamic parametric analysis, and this function also fails to take the specificity of the IF into consideration [11,13,14]. Tao et al. recently proposed a novel protocol that involves a dual-injection procedure to acquire a patient-specific alternate IF [18]. This effectively solves the challenges of loss of IF specificity and IF error; however, the complexity of the clinical procedure and medication costs are increased. With a dualscan protocol, IF can be derived through nonlinear fitting of data from segmented dynamic data, but the inconvenience to the patient in getting in and out of the bed and the problem of image registration seriously limit the feasibility of such a scheme.
There are two challenges with dynamic analysis based on only early-time dynamic data: a) when applying nonlinear regression (NLR) to solve a compartmental model, only some of the glucose transport parameters can be accurately calculated, and many parameters acquired from such schemes cannot coincide with that yielded from a standard 1-hour dynamic scan [19][20][21]; and b) for Ki parametric imaging using the Patlak model [22], the requirement for the linear section of the Patlak plot limits the extent of that scan duration can be shortened, and the Ki images obtained from Patlak data acquired from shortened scans will naturally suffer from serious spatial noise [20].
The superior imaging sensitivity afforded by the most advanced single-bed totalbody PET/CT system, uEXPLORER (United Imaging Healthcare Inc.), significantly facilitates investigations on shortening the scan durations of every single frame and the total dynamic series [23]. Many of the aforementioned studies are based on dynamic data from this imaging system. uEXPLORER dramatically promotes the improvement of PET imaging techniques; however, the data processing task becomes heavier with an increased imaging axial field of view (FOV). Therefore, for dynamic PET imaging, shortening the scan duration is becoming more meaningful from the aspect of reducing imaging computing costs. In this work, we investigated the possibility and feasibility of reducing the dynamic scan duration of 18 F-FDG PET with Patlak analysis and Ki parametric imaging. To address the problem of Ki imaging with a shortened scanning protocol, a data processing method proposed in our previous work was employed to denoise the Patlak data used for Ki fitting. The accuracy of the Ki values of the tumor lesions was statistically analyzed to evaluate the performance of this Patlak data processing algorithm for Ki parametric imaging.

Total-body dynamic PET data
The dynamic PET imaging data of 15 oncological patients acquired with uEXPLORER using 18 F-FDG were used for the Ki imaging experiments in this study. These 15 patients had different types of tumors. The dynamic scans lasted for 1 hour immediately after an intravenous injection of 18 F-FDG at different dose levels. The details of these dynamic PET data, such as the clinical diagnosis reports, patientspecific injection dose and transverse and coronal PET images indicating the representative tumor lesions, in the supplementary document. The corrected dynamic projection data were divided into 25 frames (30 s×1, 10 s×3, 30 s×4, 60 s×5, 180 s×4 and 300 s×8). In this temporal dividing protocol, the 19 th frame precisely represents the last frame of the first 30 min of data. The dynamic images were reconstructed via 3D time-of-flight (TOF) list-mode ordered-subsets expectation maximization (OS-EM) with 3 iterations and 20 subsets. The images were reconstructed into 192 × 192 × 673 matrices with a FOV of 600 mm and a slice thickness of 2.886 mm.

Patlak Ki parametric imaging
In our study, the parametric images were calculated using the Patlak linear graphical model [22] based on the effective time frames, which were defined as the frames corresponding to the monotonously descending section of the extracted input function. The applied Patlak model is shown in Eq. 1: where ( ) is the measured time-activity curve (TAC) of a voxel, and is the reference time point corresponding to the first effective frame. The input function ( ) was extracted from a 4 × 4 × 4 volume of interest (VOI) in the thoracic aorta of every dPET image series. The Ki image was obtained via linear fitting based on the measured ( ) and ( ). The flowchart for Ki parametric imaging in this study is shown in Figure 1. As shown, the calculated Patlak data underwent denoising processing before being applied for Ki parametric fitting.
The scanning temporal protocols for parametric imaging in this study are shown in Figure 2. Patlak data that corresponded to the first 30 min PI were used for Ki parametric imaging fitting. The Ki images obtained from the integrated data of a 60min scan acted were used at the reference images for evaluation.

Patlak data processing algorithm
The denoising method proposed in our previous work for dynamic PET image postprocessing was modified and used as the data processing algorithm for the Patlak data in this study. In our previous work, this method denoised the dynamic PET image series via the correction of voxel-level TACs [24], while in this work, the corrections were applied for voxel-level Patlak data. The pixel-level Patlak curve, ( ), can be similarly defined by Eq. (1). The independent variable and the dependent variable can be written by and Figure 3 showed the flowchart of the Patlak data processing method. The Patlak curve denoising operations based on third-order Hermite interpolation (THI) [25] and the data fidelities based on the original noisy curves are the two sections that are iteratively and alternatively carried out in the main loop of the algorithm. Every THI denoising step (including the initial THI denoising step) is a subloop in which the Patlak curve under processing is updated several times (25 iterations for the THI subloop are used in this study). The curve updating can be represented by where denotes the updating epoch, is an updating relaxation parameter ( = 0.3 is applied in this study), and � − is a re-estimated curve from based on THI theory. � − is determined by In (5), is the number of parallel interpolations ( = 20 is applied in this study), and � , is the ℎ individually interpolated curve based on randomly selected reference points from , namely, where is the curve index set of reference points for the ℎ individual interpolation.
� , , � represents the function conducting individual THI, for which and are the reference data coordinates and the corresponding data values, respectively, and are the coordinates of the target data points. This function was formulated by Akima [25] and is described in detail in Appendix I.
The fidelity of the denoised Patlak curve yielded by the last iteration in the main loop, * , ( ) , consists of two parallel processes: up-fidelity and bottom-fidelity for data points that are smaller and larger than the corresponding original data points, respectively. Specifically, for up-fidelity, an overshoot parameter and spatiotemporalspecific updating weights are involved. The details of the fidelity operation are given in Appendix II.

Effect of Patlak Data Processing
Original and processed voxel-level Patlak plots extracted from three types of tissues, namely, tumor, myocardium and the lung, are compared in Figure 4 for patients #9 and #12. Noisy Patlak curves are effectively smoothed without curve tendency distortion by the proposed data Patlak data processing method, showing that this processing is nearly equivalent to finding a fitted function from the given Patlak curve. Note that the linear fittings that yields the Ki values is carried out based on the last four data points in these plots, which correspond to imaging data acquired from approximately 14 to 30 min PI. For data in this section, the processed Patlak curves have higher linearity than the original curves, which may help illustrate how the proposed algorithm works in generating better Ki parametric images with this shortened scan duration. Figure 5 shows the transverse Ki parametric imaging results from the original and processed Patlak data obtained with a shortened scan duration and the reference data calculated with a standard scan time of 1 hour. The results shown are based on the dynamic data of patients #1, #2, #4, #10 and #25. Visually, Ki images corresponding to the original data contain a relatively high level of noise, but such noise is effectively eliminated by the proposed data processing method. As shown by the subfigures, the tumor structures were effectively restored by our Patlak data processing method. In addition, the effect of noise reduction is more emphasized in the nontumor region, which brings better image structure restoration and global Ki imaging quality. The comparisons of the coronal Ki images of patients #5, #7, #12 and #14 in Figure 6 illustrate a similar result: the processed Ki images have better performance in noise reduction and tissue structure restoration.

Accuracy of Tumor Lesion Ki values
To explore the accuracy of the resultant Ki values from the Patlak data processing in this study, regions of interest (ROIs) with sizes of 3×3×4 or 4×4×4 (depending on how easily the ROI could be placed) were extracted from a representative tumor lesion in the Ki image of each patient's data, and box plots indicating the distributions of Ki values in these ROIs are presented in Figure 7. The results show that, for most cases, under the condition of a 30-min dynamic scan, the distribution of Ki values resulting from the processed Patlak data are more coincident with that given by the reference data of the 60-min dynamic scan than with that from the non-processed original data. This reveals that the proposed Patlak data processing method yields better Ki imaging quality by providing more accurate voxel-level Ki values, without over-smoothing concerns.

Discussion
Ki parametric images obtained from a shortened scanning protocol have a low signal-to-noise ratio (SNR), and the consequent image structure distortion seriously restrict their function in clinical diagnosis. From this aspect, the results given in Figure  4-6 illustrate the crucial role of the proposed Patlak data processing algorithm in the task of Ki imaging. By eliminating signal errors of the Patlak data in the temporal domain, the noise level in the spatial domain of the resulting Ki images can also be effectively controlled. This results in a higher Ki image SNR and consequently better structure restoration for tumor lesions and surrounding tissues with a low metabolic rate. The advantage of tumor lesion contrast enhancement for Ki parametric imaging may be undermined by the imaging error from shortened dynamic scan protocol, which further illustrates the benefits afforded by this applied data processing method from this aspect. Boxplots of the data of patients #6, #8 and #13 demonstrate the effect of contrast restoration with the proposed imaging method: although the response of the tumor is flattened by imaging error, the proposed data processing method restores the response as much as possible. Figure 8 gives a direct visual comparison between the processed Ki images obtained with shortened scan duration and the PET images (the last frame of the dynamic series) conventionally used in clinical oncological diagnosis. The tumor lesions in the 30-min Ki images are remarkable and sufficiently clear, revealing the possibility that the Ki images calculated from shortened dynamic PET scans can be applied to clinical diagnosis, possibly replacing conventional static SUV images. This leads to great benefits, as the regular duration for clinical static PET scans can be shortened to 30 mins from nearly 60 mins PI. The Ki value measures the metabolic rate of glucose in human tissue, and the difference in temporal tendencies of glucose metabolism between tumors and normal tissues in the first 30 mins is sufficient for tumor lesion detection.
In this work, for Ki imaging using dPET data corresponding to only the first 30 mins PI, only four data points corresponding to 14-30 mins PI in the processed Patlak plots were used for Ki value fitting. To explore whether the temporal resolution of dynamic imaging influences the performance of Ki imaging under this condition of shortened scan duration, we carried out parametric imaging based on dPET datasets included in this study (patients #1 and #11) with higher temporal resolution (HTR). In the HTR series, dPET images of the first 30 mins PI were reconstructed into 30 frames (30 s× 1, 10 s× 3, 30 s× 4, 60 s× 17, 120 s× 5 and 300 s× 8) with the same method described in Section 2. As shown in Figure 9, the performance of Ki imaging with the HTR images is obviously inferior to that of parametric imaging with the normal dPET series. This may be because the temporal noise from the increased imaging temporal resolution significantly affects the measurement of temporal tendency in the first 30 mins. Therefore, a moderate temporal resolution for dPET imaging is recommended when implementing the proposed Ki imaging method, which is also desirable from the aspect of computational efficiency for dPET reconstruction and data processing.

Conclusion
In this study, we conducted total-body Ki parametric analysis based on Patlak model of dynamic 18 F-FDG PET data obtained with a shortened scan duration (30 mins). Our work demonstrates that with the help of the Patlak data processing algorithm, Ki images obtained with this shortened imaging temporal protocol are satisfactory in terms of image noise reduction, tissue structure restoration and accuracy of Ki estimations. Shortening the scan duration of 18 F-FDG dynamic PET with a total-body PET imaging system (uEXPLORER) may lead to great benefits for the promotion of dynamic PET imaging and is both possible and feasible based on our proposed Patlak data processing algorithm.