A Virtual Clinical Trial to Detect Changes in Tumor Uptake with PET using Lesion Embedding

Abstract

We conducted a virtual clinical trial (VCT) using patient data from the PennPET Explorer scanner and embedded lesions to model tumor response to therapy as measured by a change in the standardized uptake value (SUV). Two patient data sets (high and low BMI) were bootstrapped prior to embedding data from separately acquired sphere-in-air data sets. In each bootstrapped dataset, 20 small lesions (1 cm diameter) were embedded, 10 each in lung and liver organs. Multiple scans were reconstructed using list-mode time of flight-ordered subset expectation maximization (TOF-OSEM) for varying scan durations and pre-defined lesion uptake values. The resulting SUV measurements were utilized to construct receiver operating characteristic (ROC) curves to evaluate the system’s ability to discriminate between different values of lesion uptake corresponding to therapy-induced changes. The area under the ROC curve (AUC) was used as a summary metric of this discrimination performance. Our results demonstrate that long axial field of view (LAFOV) PET scanners with high sensitivity may effectively detect early tumor response to therapy, even with brief scan durations (1 minute scan) or, equivalently, with reduced activity. The AUC values change very slowly with increasing scan duration, suggesting that performance is not primarily limited by count statistics within the studied range. The AUC values are also not very sensitive to the patient BMI or the type of SUV metrics (i.e. SUV_mean and SUV_max), and are relatively independent of the local background (organ) uptake. The primary factor determining the AUC values seems to be the absolute change in lesion uptake that will be sensitive to partial volume effects as determined by the scanner spatial resolution. Hence, a standard axial field-of-view scanner with similar spatial resolution will likely perform as well as a LAFOV scanner after appropriate compensation for the sensitivity differences.

I. Introduction

PET/CT imaging plays an important clinical role as an imaging tool for diagnosis, staging, and monitoring of therapy response in cancer patients [1–5]. While there have been efforts to harmonize image reconstruction in order to achieve equivalent image quality over different scanners [6–10], limited work has focused on how to evaluate the impact on quantitative performance due to injected activity and imaging time. Clinical scan protocols have traditionally been based on the noise equivalent count (NEC) metric [11–13]. In practice, however, dose and time settings vary across sites due to factors such as patient throughput, scanner availability, and resource allocation (e.g., number of uptake rooms, scanner age, and overall workflow efficiency). This variability highlights a critical knowledge gap: the lack of quantitative, task-based evaluation of scan time and dose trade-offs in PET response assessment.

To address this gap, we develop a virtual clinical trial (VCT) framework, using patient data together with lesion embedding methodology [14–17] to model the task of detecting changes in tumor uptake following therapy. This approach allows controlled evaluation of the quantitative task of detecting therapy-induced changes in tumor uptake, using receiver operating characteristic (ROC) analysis to assess performance. The recent advancements in PET technology include a new generation of scanners with time-of-flight (TOF) resolution below 250 ps [18–23], which have the potential to enhance image quality and quantitative performance across a wide range of clinical applications. In parallel, long axial field-of-view (LAFOV) PET systems [20–24] offer significantly higher effective sensitivity, enabling novel acquisition protocols and advanced imaging strategies. A task-based evaluation is therefore critical to fully leverage the advantages offered by these technological improvements, particularly in the context of optimizing clinical performance and expanding research applications. For our studies we use the PennPET Explorer [22, 23] that offers both excellent TOF and a long AFOV. The evaluation is based on methodology used in [25], where the authors have used Monte Carlo simulations of an anthropomorphic phantom to study the impact of various sources of measurement variability on response assessment. In contrast, our work utilizes patient data and studies the advantage of advanced PET instrumentation in enhancing performance and optimizing the imaging time based on a clinically relevant task of measuring tumor response at different stages of therapy.

II. Materials and Methods

We used FDG data sets from two disease-free subjects acquired on the PennPET Explorer [22, 23] (142 cm AFOV, 250 ps TOF resolution), with BMI of 21 (small patient) and 36 (large patient). The injected activity was 9.3 and 11.8 mCi for the small and large subject, respectively. The final 3 minutes of data were extracted from the 60-min dynamic data set for our studies. The data were obtained from a pilot study [26], approved by the University of Pennsylvania Institutional Review Board, to investigate the pharmacokinetic effects of nutritional ketone ester on brain ketone and glucose metabolism in alcohol use disorder. Data from this study were used since the whole-body FDG tracer distribution in the liver and lung tissue represents that of a disease-free individual and therefore serves as the normal background for lesion embedding. To enable the assessment of quantitative metrics like measured uptake, contrast recovery and detectability, we employed a synthetic lesion embedding approach [14–17]. This method integrates independently measured “sphere-in-air” data into patient list-mode data, following the procedure described by Daube-Witherspoon et al. [16]. Briefly, list-mode data were first acquired from 1 cm diameter spheres containing known activity, positioned on a predefined grid to span the liver and lung volumes of a human subject within the scanner. For embedding, the number of sphere events to be added was first determined from the desired lesion-to-background ratio (LBR), the local background activity concentration as measured in the patient image, and known lesion volume. The sphere list-mode events are then statistically attenuated using the patient transmission prior to statistically merging them with the patient list-mode events, thereby generating a new list file of patient data with embedded lesions. To enhance the statistical reliability of pre- and post-therapy uptake measurements and account for background variability, bootstrapping of this data was also performed at the event level. Each patient data set (list file) was resampled ten times prior to lesion embedding implementation. Exclusively twenty, 1 cm diameter spheres, acquired separately in air, were then embedded in the liver and lung regions (ten in each) of each bootstrapped patient data set, representing clinical cases of cancer metastasis (see Fig.1). For this work, the initial pre-therapy lesion uptake was fixed at 7.6:1 relative to the organ background, consistent with our prior study showing good detectability for 1 cm lung lesions after a 60-second scan [17]. Based on this ratio, the resulting pre-therapy standardized uptake values (SUVs) were 3.7 in the lung and 12.2 in the liver for the patient with BMI 21, and 2.4 in the lung and 7.3 in the liver for the patient with BMI 36. Change in lesion uptake, post-therapy, was modeled as being ±10% and ±25%, leading to 50 data sets per patient (10 bootstraps and 5 lesion uptake values). Data were reconstructed using list-mode TOF ordered subset expectation maximization (TOF-OSEM) (5 iterations × 25 subsets) [27] for six different imaging times of 180, 120, 60, 30, 15, and 7.5 seconds. Table I summarizes key simulation parameters used in this study.

Fig. 1.

Transverse slice from a reconstructed image with a 1 cm diameter embedded sphere in the liver (a) and lung (b). The images are from a 3-minute FDG scan on the PennPET Explorer scanner.

Table I.

Summary of simulation parameters used in lesion embedding and reconstruction

Parameter	Small patient (BMI = 21)	Large patient (BMI = 36)
Injected activity (mCi)	9.3	11.8
Pre-therapy lesion-to-background ratio (LBR)	7.6:1	7.6:1
Pre-therapy lesion	3.7	2.4
SUV (lung)
Pre-therapy lesion	12.2	7.3
SUV (liver)
Modeled therapy response	±10%, ±25% change in uptake	±10%, ±25% change in uptake
Lesion size	1 cm diameter	1 cm diameter
Number of lesions per dataset	20 (10 lung, 10 liver)	20 (10 lung, 10 liver)
Bootstrapped list data per patient	10 (event-level)	10 (event-level)
Imaging times evaluated	180, 120, 60, 30, 15, 7.5 s	180, 120, 60, 30, 15, 7.5 s

The list mode algorithm includes optimized basis functions [28] to suppress image noise while preserving signal; hence, no post filtering was used. For image analysis, measured uptake (standardized uptake value, or SUV) was quantified as the mean SUV within a 1 cm diameter spherical volume of interest (VOI) centered on the known lesion location. Therefore, 100 measured SUVs per patient/organ/uptake/scan time were generated. Besides the SUV_mean (average SUV within the VOI), the SUV_max (maximum SUV value within the VOI) metric was also considered for tumor uptake measurement. Histograms of the pre- and post-therapy SUV measures were then created from the distribution of SUVs per patient/organ/scan time. Subsequently, ROC curves [29] were constructed by thresholding the measured lesion uptake values (SUV_mean and SUV_max) to discriminate between pre- and post-therapy distributions. Thresholds were defined over the full observed SUV range, incremented in steps of 0.1 SUV, to ensure consistent sampling across patient size, organ, and scan duration. For each threshold, true and false positive fractions were computed and assembled into the ROC curve. For each ROC curve, the area under the curve (AUC) was then calculated as the figure of merit, representing the probability of correctly discriminating the modeled change in tumor uptake. To quantify uncertainty in AUC estimation, we applied a bootstrap procedure: each set of pre- and post-therapy histograms was resampled 100 times with replacement, and the corresponding AUC value was calculated for each resample. The standard deviation across these bootstrapped AUC values was reported as the uncertainty of each AUC value. The whole procedure of the AUC calculation for this study is shown in Fig. 2.

Fig. 2.

Schematic workflow for AUC calculation for a study with a pre-therapy Uptake (j) and post-therapy Uptake (k) as shown. The ΔSUV refers to change in SUV in the flowchart.

III. Results

Fig. 3 shows a sample histogram for pre- and post-therapy SUV_mean’s used to create an ROC curve for one patient study between pre- and post-therapy measurements as function of SUV_mean threshold, under a given set of imaging conditions. The corresponding ROC curves obtained from the previous patient study as a function of scan duration is shown in Fig. 4. As one would expect, with increased scan time, the ROC performance improves.

Fig. 3.

Pre-therapy and post-therapy SUV_mean histograms for a patient with BMI of 21 and a −25% change in tumor uptake at 3-minute scan in liver organ.

Fig. 4.

ROC curves for change in uptake of −25% in a patient with BMI of 21 for varying scan durations.

Fig. 5 plots the AUC value as a function of the two different measured SUV metrics (SUV_mean and SUV_max) for change in uptake (Δuptake) of ±25% in the lung and liver organs in two patients for 180 seconds scan time. The results indicate that AUCs for both SUV metrics are comparable (taking the error bars into account). Note that we use SUV_mean as the metric for uptake measurements for all the subsequent results.

Fig. 5.

AUC values plotted as a function of a ±25% change in lesion uptake. Results are shown for two different patient BMIs (21 and 36), lesions present in lung (top) or liver (bottom), and SUV measured as SUV_mean and SUV_max. Scan time was fixed at 180 seconds.

Fig. 6 shows the AUC values as a function of Δuptake for lesions present in the lung and liver for the two patient BMIs. Higher AUC values are achieved with larger Δuptake as expected. The AUC values are higher for liver lesions when compared to equivalent lung lesion primarily due to the fact that the absolute uptake in the liver lesions was higher since we modeled a fixed 7.6:1 uptake relative to organ’s uptake. Furthermore, the results show that the AUC values do not change substantially with patient BMI.

Fig. 6.

AUC values for lung and liver lesions in two patients for varying change in uptake but fixed scan time (180 seconds). The measured SUV metric was SUV_mean.

Comparing the AUCs for a decrease in uptake to an equivalent increase, we find slightly higher AUCs when there is a decrease in the uptake value, indicating slightly better performance for detecting a positive response to therapy. Benefiting from the very high sensitivity of a LAFOV scanner such as the PennPET Explorer, one can achieve a high AUC value of ≥ 0.80 for Δuptake of ±25%, although the AUC values decrease when looking for a smaller change in tumor uptake post-therapy (±10%) using 180 second scans.

In Fig. 7 we plot AUC values as a function of scan time. As can be seen, reduced scan times show a small but systematic drop in AUC values, for example for Δuptake of −25% in a lung lesion the AUC changes from ≥ 0.80 to ≥ 0.70 when reducing scan time from 180 seconds to as low as 7.5 seconds. In Fig. 7, we also see that the AUC values for a ±25% Δuptake in the lung lesion (see the blue bars in the top row) is comparable to a ±10% Δuptake in the liver lesion (see red bars in the bottom row). This is likely because in this study we used a fixed lesion-to-background uptake ratio during lesion embedding, resulting in different absolute lesion uptake in each organ (higher in liver). In particular, a ±25% change in lung lesion uptake is similar in count levels to a ±10% change in liver lesion uptake, indicating that AUC performance is driven by the absolute change in uptake. This equivalence highlights that AUC values are more sensitive to the absolute change in lesion uptake. Interestingly, we also observe that a 60-second scan is sufficient to detect an Δuptake of ±25% in lung lesions with AUC values ranging from 0.75 to 0.80. For smaller uptake changes (Δuptake of ±10%), the AUC decreases to approximately 0.60–0.65 for the same scan duration. Notably, extending the scan time beyond 60 seconds yields only marginal improvements in AUC, suggesting that statistical noise is not the primary limiting factor in detecting these uptake changes. This further supports the idea that for our study the absolute change in lesion uptake which is related to the measured SUV — rather than scan duration alone — drives the ability to detect changes in uptake, and that similar performance may be achieved on lower-sensitivity systems (e.g., standard axial field-of-view (SAFOV) PET) with appropriately scaled scan durations.

Fig. 7.

AUC values shown as a function of scan time for the two patient BMIs with lesions present in liver and lung. Results are shown for shown separately for varying change in uptake. The measured SUV metric was SUV_mean.

IV. DISCUSSION

In this study, we developed a VCT framework using patient data acquired on the PennPET Explorer to evaluate the ability of the system to detect changes in tumor uptake measurements after therapy. The focus was to assess the impact of scanner sensitivity and scan duration on the reliable detection of changes in tumor uptake, as reflected by ROC-based performance metrics. Our results demonstrate that a LAFOV scanner with 250 ps TOF resolution enables good detection of a ±25% change in tumor uptake (AUC ≥ 0.80) for scan durations as low as 1 minute. A 1–2 minute scan on LAFOV PET has been shown to be comparable to a 10–16 minutes scan in a conventional whole-body PET scanner with similar TOF capability [30, 31]. Since a short scan time can also be equated as reduced injected activity, our results demonstrate that in a LAFOV PET scanner we can reliably measure a ±25% change in tumor uptake with lower injected dose to the patient, making serial scans post-therapy more practical.

Our conclusions are also relatively insensitive to the patient size (range BMI 21–36) and two types of SUV measures (SUV_mean and SUV_max) that we used in our work. This is particularly important in clinical settings, as it suggests that the response prediction performance of the scanner remains robust across a wide range of patient body compositions. In clinical practice, both SUV_mean and SUV_max measures have strengths and limitations. SUV_mean is generally more robust to statistical noise, making it particularly reliable for low-count or short-duration scans, whereas SUV_max can be more sensitive to local hotspots or outlier voxels, which may lead to higher variability under noisy conditions. Therefore, while SUV_max can highlight the most intense region of uptake, SUV_mean provides a more stable measure of average lesion activity, which is advantageous when evaluating small lesions or performing serial post-therapy comparisons with reduced scan time or injected activity. While a comparison between SUV_mean and SUV_max was performed for a representative case—demonstrating comparable AUC values within error margins—all subsequent results presented in this work are based on SUV_mean. This choice reflects the practical advantages of using SUV_mean for well-defined spherical lesions and is consistent throughout the study. Comparing performance across the two organs studied here (lung and liver), we find that the AUC values are systematically higher in the liver compared to the lung, indicating better performance for detecting response in liver lesions. However, due to a lower pre-therapy uptake in the lung, a 25% change in the lung tumor uptake corresponds to a 10% change in the liver tumor uptake when considering the absolute change in uptake. Our results in Fig. 7 indicate that the AUC values for ±25% Δuptake in the lung are similar to ±10% Δuptake in the liver, indicating that the AUC results are not dependent on the organ (and its background uptake), but instead are more a function of the absolute change in lesion uptake that is related to the measured SUV.

The gradual change in AUC values with increasing scan duration also suggests that performance is not primarily limited by count statistics within the studied range. Instead, these results show that the performance is likely determined by the scanner spatial resolution that determines the level of partial volume effect present in the image, and hence, defines our ability to accurately measure a change in uptake. A small change in uptake will be more difficult to detect in scanners with poor spatial resolution due to increased partial volume effect particularly for small lesions. For example, a 1 cm spherical lesion on the PennPET Explorer exhibits an expected contrast recovery of approximately 50–55 % [23], meaning that measured SUVs are reduced by nearly half compared to true uptake. This effectively diminishes the detectability of small uptake changes—for instance, a true 25% change in uptake would be observed as only a 12.5–13.75% change—thereby limiting sensitivity to small uptake differences. The challenge is compounded in heterogeneous background organs such as the lung, where local variations in SUV can be of the same order as these reduced uptake differences. In such settings, distinguishing a modest true biological change from normal background variability becomes increasingly difficult. These combined effects highlight that sensitivity to small SUV changes is constrained not only by partial volume averaging but also by background heterogeneity, both of which must be considered in interpreting quantitative response metrics. However, it should also be noted that we used a relatively small sphere to define our tumor of interest (1 cm diameter), whereas clinically one is likely to assess larger tumors that are less affected by the limited scanner spatial resolution. This trend implies that comparable AUC values could be achieved with longer scan times on scanners with lower sensitivity but similar spatial resolution. For example, LAFOV PET scanners—such as the PennPET Explorer—offer total-body sensitivity gains of 10–20 × compared to standard axial field-of-view (SAFOV) systems, or approximately 2–3 × higher sensitivity when localized to specific organs [32]. As such, the uptake measurement performance demonstrated in this study with short-duration scans on a LAFOV system may be extrapolated to longer scan durations on SAFOV systems, making these findings broadly applicable across different PET system configurations.

A key limitation of this study is that the task investigated was restricted to the discrimination of pre-defined changes in tumor uptake, as quantified using ROC-based metrics. While this framework is well suited for assessing the ability of a PET system to distinguish between different uptake-change classes, it does not address the more clinically relevant problem of quantitatively estimating the magnitude of uptake change on a per-lesion basis. Accurate estimation of the absolute change in SUV, rather than simple detection of change, is critical for treatment response assessment and clinical decision-making, but was beyond the scope of the present work. Extending this framework to directly evaluate the precision and bias of uptake change estimation represents an important direction for future research.

A second limitation of this study was its use of 1 cm spherical lesions and a fixed lesion-to-background ratio for all evaluations, thereby not capturing the full diversity of lesion sizes, shapes, and uptake contrasts observed clinically. In the future one could in principle extend this work to better represent the full clinical diversity of lesion contrasts, shapes, and sizes. However, our results indicate that the AUC values are similar for same change in lesion uptake, indicating that at least the pre-therapy lesion contrast by itself will likely not affect the AUC performance. A third limitation of our study is that the lesion embedding methodology does not include the impact of scattered events from the sphere. However, as shown in our past work [33], this effect was not significant, at least for the sphere size used in this work.

Future work will focus on extending the methodology to evaluate irregularly shaped lesions that better reflect clinical tumor morphology and influence uptake measurement accuracy. Incorporating a broader range of lesion sizes and uptake contrasts, as well as validating the framework with real tumor datasets such as those available through The Cancer Imaging Archive, will be critical for establishing clinical relevance. In addition, AI-based classification approaches [34] could be integrated to augment lesion detection and improve sensitivity to subtle uptake changes, thereby enhancing the utility of VCT studies for therapy response assessment.

V. CONCLUSIONS

The results of this study demonstrate the ability of LAFOV scanners with excellent TOF resolution in measuring tumor response to therapy. Our findings show that the PennPET Explorer can reliably measure ±25% changes in tumor uptake for short 1 minute scan duration. The robustness of AUC values across BMI and two different SUV measures for spherical lesions further support the clinical utility of this approach, particularly in scenarios involving well-defined tumor shapes. The ability to accurately measure a change in lesion uptake is relatively independent of the local background (organ) uptake and is determined more by the absolute change in lesion uptake.