Dynamic whole-body PET (dWB-PET) imaging provides additional information to usual static SUV metrics, particularly on the tracer influx rate (Ki) through pathological or physiological tissues [13]. Knowledge of spatio-temporal distribution of radiotracer could allow to better differentiate tumor lesions from inflammation and thus influence the response assessment to therapy in oncology. Although the principle of multiparametric PET imaging has been described for several decades, its use in clinical routine has only become feasible in rare nuclear medicine departments, equipped with specific PET systems integrating automated workflows. Indeed, such workflows are simpler and less invasive than arterial blood sampling, in using an image-derived input function (IDIF) to quantify blood activity concentration through acquisition time but are still time-consuming. Nevertheless, optimizations are still needed to make dWB-PET imaging acceptable in daily practice. The recent proposal for population-based IF (PBIF) modeling is an area of research allowing to significantly reduce imaging time and improve patient comfort [31,32,33].

Our study shows the clinical feasibility of a short duration WB dynamic 18FDG-PET imaging using a PBIF without significant differences in Ki quantification, in a cohort of patients with metastatic melanoma treated by immunotherapy.

In our study, the PBIF model was created from a cohort of 17 controlled patients, a sufficient sample size according to published data (varying from 11 to 23) [28, 40]. The low mean relative bias (0.5%) between the AUCs of the scaled PBIF and the IDIFs of the MM cohort shows that a model created from healthy patients can be applied to a specific pathology.

We choose to collect for each case an IDIF using an automatically defined volume of interest (VOI) in the left ventricle. In a previous study of 24 patients, Sari et al. [41] found a good agreement between the amplitudes of the peaks and tails of IDIFs derived from ascending aorta, descending aorta, left ventricle, and left atrium. Most importantly, they emphasized that an IDIF measured from the carotid arteries underestimated the AUC due to uncorrected partial volume effects.

After resampling of different TACs, we synchronize IDIF on the shortest time-to-peaks, unlike Sari et al. [33] who used the mean time-to-peak. The use of the shortest time to peak could be one of the causes of the negative bias observed when using PBIF. Indeed, when using the model, the real time-to-peak information cannot be taken into account. The impact of such difference on our overall results will be assessed in a further comparative study.

We fitted the average IDIF according to the Feng model. This mathematical function was the most used in the recent literature, as example by Naganawa et al., Dias et al. and Sari et al. [28, 33, 42]. Like their results, we found an excellent correlation between the measurements and the model (R2 = 0.998). Our maximum shape deviation was observed only in the first 20 s post-injection and was quantified at an insignificant fraction of 0.5% of the total AUC. This result was comparable to those reported in the respective series of Naganawa et al. [28] (− 1 ± 6% for a 15–45-min scaling time window, and 3 ± 6% for a 30–60 scaling time window) and Dias et al. [32] (− 3 ± 6% for a 30–50 scaling time window; − 2 ± 7% for a 40–60 scaling time window; both with Feng model but compared to arterial IF). The fit of the input function is an important factor in the quality of the results. Overfitting or underfitting can lead to over- or underestimation of Ki. Studies have shown that an error of 20% on the AUC of the IF leads to a deviation of around 4% on the Ki [29, 43]. Other models can be used to describe IF. We used a sum of a gamma variate function with three-exponential method, while Dias et al. [32] also compared with a four-exponential model with good results (2% to 0% bias, respectively, for a 40–60 scaling time window) but comparing PBIF to AIF.

One of the parameters that will influence parametric reconstruction is the time window used to scale the PBIF, especially when using the latest points [28]. In our study, PBIF was scaled using the whole late dynamic acquisition (i.e., 11–70 min). It would have been of interest to try multiple scaling windows as already assessed in the literature [31, 41]. Nevertheless, this was not possible in our study due to the software limitations of our PET system. However, this concern does not appear to be crucial for the PBIF scaling. Indeed, Sari et al. [33] evaluated different scaling windows and concluded that the use of a late window (i.e., 55–65 min) would provide the lowest bias in their study. Thus, our results remain valid because the use of a late scaling window, concomitant with the window used for the Patlak analysis, will provide a similar accuracy.

Once the IF has been reconstructed, our method of testing 9 different time windows (TWs) of dWB-PET data (3 or 4 passes at 15- to 70-min post-injection time) to compare scaled PBIF and measured IDIF on the respective Patlak reconstructions is consistent with the literature [28, 31, 32]. In our results, the TW used to perform the direct 4D nested Patlak reconstruction affected the Ki measurements, even if the bias was overall acceptable (less than 10% regardless of the time window). Indeed, the Ki bias was lower in late time windows (around 5% for 35–53 min, 45–61 min, 45–70 min time windows) than in early time windows (8% to 9% for 20–35 min, 20–45 min, 20–61 min time windows).

Our best selected reconstructions in terms of optimal use in routine practice and statistical results were 45–61 min and 45–70 min TWs. Indeed, we showed a very good correlation between IDIF and PBIF reconstructions (bias − 5.2% and − 4.9%, SD 4.1% and 6.5%, R2 0.999 and 0.997, respectively). We observed a slight underestimating of Ki metrics, especially in case of high value as already described [44]. Effectively the 5 highest Ki mean values showing higher differences between IDIF5_7 and PBIF5_7 (− 0.8 to − 1.4 absolute difference, − 7 to − 15%) were measured on the same patient showing very high tumor uptake on standard PET images (SUV values around 100).

Surprisingly, we found a negative bias, although it has been described between + 1.5 [33] to + 7.4% [31] and even up to + 23% [31] but with a very short 10-min acquisition. It can also be a PET system effect. Indeed, the positive bias occurs on large FOV PET system, while on standard FOV the bias remains negative [28, 32]. In any cases, the bias generated by our model remained lower than the difference between pathological and physiological Ki values. These results are promising, as a 20 min of WBdyn PET acquisition time in daily practice remains largely acceptable for both the patient’s compliance and the department’s organization.

We chose to use a VOI predefined on the gold-standard 3D SUV image for Ki metrics extraction of each pathological lesion or physiological organ in the IDIF and PBIF parametric images. Although parametric PET images are known to show a better contrast and attain superior quantification, they are also typically more sensitive to noise [37] and may suffer from kinetic artifacts compared to static SUV images. Indeed, SUV images use projection acquired 1h post-injection (PI), while 2_4 Patlak reconstruction use projection acquired during the 20–35 min PI for example. Consequently, we decided to measure both mean and maximum Ki metrics in VOIs because mean values are less sensitive to noise, while maximum values are less sensitive to kinetic blur. Ultimately, both measurements showed similar bias (difference lower than 1%) and correlations when comparing IDIF and PBIF-based reconstructions, regardless of the chosen time window.

We found no significant difference in Ki values of both tumoral lesion, inflammatory disease, and physiological uptakes in comparing IDIF5_7 and PBIF5_7. Regarding physiological Ki values, our results were comparable with literature data [42, 45]. As Dias et al., we found widest variation was found on myocardial Ki values (range from 0.54 to 8.55 100ml/ml/min with PBIF5_7), probably explained by the absence of specific cardiac free fatty-acid consumption diet to suppress physiological myocardial uptake. Regarding tumoral lesions, no specific published data on melanoma are available for comparison. Sari et al. [45] only described 2 lesions of melanoma with mean Ki of 1.3 and 3.9 100ml/ml/min (average of 2.87 100ml/ml/min [0.83; 13.33] in our results). These findings were globally concordant, regarding Vd values, in terms of bias, less than 5%, for each time window. However, the dispersion is higher than for Ki, with SD ranging from 6.4 to 22.1%.

To the best of our knowledge, this is the first study applying PBIF approach trying to resolve a clinical issue in routine practice, in applying our model to a cohort of 20 patients with metastatic melanoma (MM) treated by immunotherapy (ICI). In these preliminary results of the IMMUNOPET2 study, we found that 4D whole-body dynamic PET images might be capable to differentiate progression disease (PD) to pseudo-progression (PP) (mean Ki values 2.87 [0.83–13.33] versus 1.13 [0.37–2.04]). This capability of Ki values to differentiate malignant to inflammatory lesions has recently been suggested by Shawran et al. [21], but in a large selection of tumor type and inflammations etiologies.

As a perspective, another axis of optimization would be to assess the impact of reconstruction parameters on Ki measurements. Indeed, Patlak imaging is sensitive to noise, though the impact is relatively less for direct 4D nested Patlak reconstructions. This would allow us to evaluate the reproducibility of the Ki measurements and to establish the best ratio between detectability and image quality. Wu et al. [37] have already published a study with this approach by applying a denoising filter during reconstructions. In a cohort of 65 patients, they proved a similar detectability of lesions by reducing acquisition time by two. Finally, a more efficient and less complex approach to manage image noise in dynamic WB PET data would be to exploit high sensitivity scanners, such as large axial FOV or total-body PET systems [31, 33, 37]. Although this type of device is not yet widely available from manufacturers, is currently expensive and thus not widely adoptable in clinic [46, 47], it will definitely open a new era in parametric PET imaging.