From RECIST to PERCIST: Evolving Considerations for PET Response Criteria in Solid Tumors

WHO Criteria

Moertel and Hanley's work and the development of a variety of promising anticancer therapies, mainly cytotoxics, in the 1960s and 1970s brought about a clear need for standardization of response criteria. Because CT of the body was not in widespread use until the early 1980s, most tumor measurements were obtained by palpation or chest radiographs. In 1979, WHO attempted to standardize treatment response assessment by publishing a handbook of criteria for solid tumor response (7). The proposed WHO methods included determining the product of the bidimensional measurement of tumors (i.e., greatest perpendicular dimensions), summing these dimensions over all tumors, and then categorizing changes in these summed products as follows: complete response—tumor has disappeared for at least 4 wk; partial response—50% or greater reduction in sum of tumor size products from baseline confirmed at 4 wk; no change—neither partial response nor complete response nor progressive disease; and progressive disease—at least a 25% increase in tumor size in one or more lesions, with no complete response, partial response, or stable disease documented before increase in size, or development of new tumor sites.

Reviewing the data of Moertel and Hanley, one would be concerned that the progressive disease category in WHO might be easy to achieve by chance changes in measurement (i.e., a 25% increase in the product of 2 measurements could occur with an approximately 11% increase in each dimension). In addition, the WHO criteria were not explicit on such factors as how many tumor foci should be measured, how small a lesion could be measured, and how progression should be defined. Thus, despite efforts at standardization, the WHO criteria did not fully standardize response assessment. The WHO criteria are still in use in some trials and are the criteria used to define clinical response rates in many trials from the past 2 decades—which are important reference studies. Although not as commonly used at present, familiarity with the WHO response criteria is essential for comparison with more recent studies using RECIST, especially as relates to the issue of when tumors progress. The WHO criteria are summarized in Table 3.

RECIST

The RESIST criteria were published in 2000 and resulted from the recognition of some limitations of the WHO criteria (8). The criteria were developed as a primary endpoint for trials assessing tumor response. In addition, between the time of development of the WHO criteria and development of RECIST, cross-sectional imaging with CT and MRI entered the practice of oncology. RECIST specified the number of target lesions to assess (up to 10), though it did not give substantial guidance on how they were to be selected, except that there should not be more than 5 per organ. RECIST assumed that transaxial imaging would be performed, most commonly with CT, and specified that only the single longest dimension of the tumor should be mentioned. Thus, RECIST implemented a unidimensional measurement of the long axis of tumors. RECIST also clearly stated that the sum of these unidimensional measurements was to be used as the metric for determining response. RECIST also specified the minimum size of the lesions to be assessed, typically 1 cm using modern CT with 5-mm or thinner slices. Lesions of adequate size for measurement are described as “measurable.” There are also designations of “target” and “nontarget” lesions (Tables 1–3⇑⇑). All target lesions are measurable. Some nontarget lesions are measurable. Both can contribute to disease progression and to complete response (Tables 1–3⇑⇑).

The RECIST categories for response include complete response—disappearance of all tumor foci for at least 4 wk; partial response—a decline of at least 30% in tumor diameters for at least 4 wk; stable disease—neither partial response nor progressive disease; and progressive disease—at least a 20% increase in the sum of all tumor diameters from the lowest tumor size. A 20% increase in tumor dimensions results in a 44% increase in the bidimensional product, substantially greater than the WHO progression criterion of 25%. One would predict progression to be later, and possibly less frequent, using RECIST than using WHO. This has been the case, and earlier progression is seen in about 7% of patients using WHO versus RECIST (8). Thus, time to disease progression can be shorter with WHO than with RECIST (for the identical patient data). When progression is due to new tumor foci (which occurs about half the time in some reports), the 2 methods would be expected to be concordant in indicating progression of disease (8). Overall, quite good concordance was seen with the 2 methods. The RECIST and WHO criteria are contrasted in Table 3.

Another consideration for anatomic and functional imaging is that many of the changes in response, from partial response to complete response, or from stable disease to partial response, are at the border zones between response groups (i.e., 48% vs. 52% change in tumor size in WHO, or 28%−32% change in RECIST (nonresponse vs. partial response, for example). These border zones are frankly quite artificial, as changes in tumor size occur on a continuum. This is why continuous, so-called waterfall, plots of fractional shrinkage or growth of tumors are becoming increasingly popular as a means of graphically displaying tumor response data (1,2,10). It is to avoid such problems that PERCIST includes providing a specific percentage reduction in the SUV (SUV lean, or SUL) from baseline, as well as noting when the information is available—the number of weeks from the start of treatment.

Therasse, Verweij, et al. recently reviewed the use of RECIST in about 60 papers and American Society of Clinical Oncology meeting abstracts (37,38). The expected delay in progression detection versus WHO was observed. In addition, recognition of challenges in certain pediatric tumors, unusually shaped tumors such as mesotheliomas, and tumors with a great deal of central necrosis or cystic changes, such as gastrointestinal stromal tumor (GIST), were noted. Overall, however, the authors believed that RECIST had been highly successful but that some improvements were needed.

RECIST 1.1

The RECIST group, which included representatives from, among others, the EORTC, the National Cancer Institute (NCI), the National Cancer Research Network, and industry, recently reported new response criteria for solid tumors, RECIST 1.1 (39). This version of RECIST, reported in January 2009, includes several updates and modifications to refine the prior RECIST criteria. Notably, RECIST 1.1 made use of a data warehouse of images and outcomes provided from a variety of clinical trials, allowing assessment of changes in tumor size based on several formulae. Although the original RECIST included size measurements of up to 10 lesions, with a maximum of 5 for any single organ; simulations in RECIST 1.1 assessed the use of 1, 2, 3, or 5 target lesions, versus the original 10. They found strong agreement in response classifications using fewer than 10 lesions, even using just 1 lesion, but even better concordance when 5 lesions were used. In randomized studies in which tumor progression is the major concern, RECIST 1.1 suggests that just 3 lesions may be used, not 5. Thus, there are potentially 50%−70% fewer tumor measurements with RECIST 1.1 than with RECIST. RECIST 1.1 also suggests that the largest lesions be used for response, as long as they are distinctly capable of being measured.

RECIST 1.1 also dealt with lymph nodes differently than did the original RECIST criteria. In the original RECIST, the longest axis of lymph nodes was to be measured and the lymph nodes had to disappear completely to secure a complete response. In RECIST 1.1, nonnodal lesions had to be 1 cm in size or larger (long axis) to be considered measurable. By contrast, in RECIST 1.1, the short axis of lymph nodes is measured; short-axis lengths greater than 1.5 cm are considered suitable for measurement, and nodes with short axes under 1 cm are considered normal. If a node disappears nearly completely and cannot be precisely measured, it is assigned a value of 5 mm. If totally absent, it becomes 0 mm. The difference between RECIST and RECIST 1.1 in lymph nodes is that the lymph node size can decline to greater than 0 and still be considered a complete response. Thus, with RECIST 1.1, especially in diseases in which lymph nodes represent a significant fraction of the total tumor burden, criteria for a complete response are less stringent than with the original RECIST. In the simulation data used in the RECIST 1.1 study, if nodal disease predominated, 23% of cases would move from partial response to complete response, whereas about 10% would move from partial response to stable disease. It should be noted that short-axis nodal diameter is added to long axis of other tumors to result in an overall tumor burden assessment in measurable lesions. This reclassification to an increased complete response rate for node-dominant disease is a major change and may be controversial as regards comparing RECIST with RECIST 1.1.

The overall definition of progressive disease also changed in RECIST 1.1 by requiring an absolute increase in the sum of the tumor dimensions of at least 5 mm. This requirement prevents a minimal (<5-mm sum of tumor long axes) 20% increase from being categorized as progressive disease. The new RECIST 1.1 criteria offer guidance on what constitutes unequivocal progression of nonmeasurable or nontarget disease. There is also a brief discussion in RECIST 1.1 of the implications of a newly positive PET scan with ¹⁸F-FDG in disease otherwise not considered to be progressing—the PET scan must be taken seriously as recurrence (39–41). Methods for classifying anatomic response in RECIST and RECIST 1.1 are detailed in Tables 1–3⇑⇑.

Although these anatomic criteria may appear to be arcane, the RECIST criteria and now, quite likely, the RECIST 1.1 criteria are or will be used in virtually every clinical trial of new solid tumor therapeutics, as response is essentially always measured. Further, regulatory agencies have accepted RECIST as the de facto standard in response assessment for clinical trials in many countries. Familiarity with the implications of trials in which response is measured using the WHO, RECIST, and RECIST 1.1 criteria is essential, as they are not identical and do not produce identical results.

Limitations of Anatomic Response Criteria

Although RECIST has been used quite extensively for the past 8 y, some concerns about the method have not been fully addressed, even in RECIST 1.1. One issue is the fundamental statistical issue of reducing intrinsically continuous data on tumor size and tumor response to a series of 4 bins of response (i.e., complete response, partial response, stable disease, and progressive disease). With such reductionism, potential valuable information that may be important is lost (1,2,4,10). For example, with some newer cancer treatments that are mainly cytostatic, longstanding stable disease is a highly beneficial outcome. Indeed, examples of such effects include the behavior of GIST tumors, in which tumor size shrinks slowly but patients live for long periods with stable disease (42,43). Similar findings of prolonged life, with limited antitumor size response by RECIST, have been seen in hepatomas treated by sorafenib (44,45). Thus, there have been attempts to use tumor characteristics other than size to assess response. For example, the Choi criteria that have been developed for GIST include assessments of the size and CT Hounsfield units of tumors before and after treatment. With the Choi criteria, a 10% decrease in size or a 15% decrease in CT Hounsfield units is associated with a good response. Although these are potentially difficult measures to make precisely, it has been generally agreed that RECIST is not adequate for GIST (42,46,47). Additional anatomic characteristics of GIST, such as the development of mural nodules, but not necessarily with tumor growth because of the predominantly cystic nature of the tumors, are indicative of progression and of a poor outcome (48,49).

Limitations of RECIST in predicting response are noted clearly in the SHARP trial, in which sorafenib, an inhibitor of vascular endothelial growth factor receptor, platelet-derived growth factor receptor, and Raf, was used in a randomized placebo-controlled trial in patients with hepatoma. In this trial of over 602 hepatoma patients who had not received previous therapy, only about 2% of the treated group and 1% of the control group had a partial response by RECIST, a figure that might lead one to conclude the drug to be inactive. However, the main endpoints of the trial were not tumor response but rather survival and progression-free survival. Because hepatomas have a bad prognosis and there is a high death rate, survival studies are feasible. At the time the study was ended, median overall survival was 10.7 mo in the sorafenib group and 7.9 mo in the placebo group (P < 0.001). The median time to radiologic progression was 5.5 mo in the sorafenib group and 2.8 mo in the placebo group (P < 0.001). Thus, clearly prolonged survival of about 3 mo was seen in this group of patients with advanced hepatocellular carcinoma treated with sorafenib, in comparison to patients treated with placebo. This substantial improvement in survival was associated with stable (not shrinking) anatomic disease (45).

In hepatomas, alternative criteria to RECIST have been developed, referred to as the EASL (European Association for the Study of the Liver) criteria (44,50). These criteria rely on contrast enhancement patterns after vascular interventional therapies and appear superior to RECIST in this limited setting. Similarly, in mesotheliomas and pediatric tumors, modifications of RECIST dealing with the peculiarities of these tumors are in place (51–53,53A).

An additional consideration for RECIST is that the most precise estimates are achieved when the same reader assesses the baseline and follow-up studies. More misclassifications and variance in response are seen when a different reader assesses the baseline and follow-up studies (54).

Tumor size is a clearly important parameter, and there is some evidence that the more rapidly a tumor shrinks, the more likely it is that the response will be durable. For example, in lymphomas, patients whose tumors shrink the most rapidly are most likely to do well, and they may need less treatment (55). Estimates of tumor volume may prove more useful than 1-dimensional methods of tumor assessment in evaluating tumor response. Caution, however, is needed even with volumes; in neoadjuvant therapy of lung cancer, early changes in lung cancer volume were shown not to be predictive of histologic response (56). Tumor histologic status was well associated with changes in tumor volumes in neoadjuvant therapy of colorectal cancer, however (57). The use of continuous as opposed to discrete sets of response has been suggested. Such continuous assessments may then lend themselves well to randomized phase II trials in which the response metrics can be compared using more standard statistical testing than concordance or κ-statistics (4).

Lymphoma

Lymphomas have had a somewhat different approach to response assessment than solid tumors. Briefly, residual or even bulky masses after therapy completion are frequent in both Hodgkin disease and non-Hodgkin lymphoma but correlate poorly with survival (58). Masses often do not regress completely after adequate (curative) treatment because of residual fibrosis and necrotic debris. The anatomic response categories of “complete remission unconfirmed” or “clinical complete remission” were created in recognition of the problem that, particularly in patients with lymphoma, anatomic response criteria often underestimate the chemotherapeutic effect (59). Patients with stable disease by conventional anatomic criteria may be cured. It has been demonstrated that adding PET to the posttherapy CT is especially useful in identifying which of these patients have achieved a satisfactory functional remission (60,61). The reader should be aware that there are well-established anatomic metrics of response in lymphoma (59). These metrics have recently been updated and modified to include PET at the end of therapy because of the limitations of anatomic imaging (Tables 4 and 5) (32,33).

Although limited in their early assessment of treatment response, and somewhat variable in terms of outcome prediction, WHO, RECIST, and RECIST 1.1 are the standard anatomic response assessments currently accepted by most regulatory agencies, and RECIST, in particular, is in widespread use in clinical trials. By contrast, it is infrequent for these response criteria to be used in routine clinical practice. Although the criteria are quite detailed, variance in response occurs because of measurement errors and the inability of anatomic processes to quickly detect functional changes in tumors resulting from early effective treatment. The delayed readouts from anatomic imaging mean that it is difficult to quickly use anatomic imaging to modify treatments in individual patients. Functional imaging with PET offers major advantages.

How Is Response Determined on PET?

Two basic approaches can be considered for assessing the metabolic changes of treatment: qualitative and quantitative. Another issue is whether a response scale should be binary (yes/no for response) or continuous (giving varying degrees of response). An additional and not fully resolved issue is whether the most metabolically active region of the tumor should be assessed or whether the entire tumor burden glycolysis and volume should be assessed. Not fully resolved, as well, is what constitutes a negative scan, a problem not unique to ¹⁸F-FDG PET (88).

Quantitative

Because PET is intrinsically a quantitative imaging method, quantitative measurement of early treatment-induced changes is an attractive potential tool for measuring subclinical response and more complete changes. The feasibility of detecting small changes in tumor glucose metabolism quantitatively was demonstrated over 15 years ago in studies of neoadjuvant treatment of primary breast cancer, for which declines in SUV of 20%−50% were seen, depending on the time from the start of treatment. These declines were evident using Ki, SUV, and the k₃ rate constant (20). More than 30 different ways to monitor tumor response have been discussed, but the SUV appears to be the most widely applied, generally correlating well with more complex analytic approaches (100,101).

The SUV is a widely used metric for assessing tissue accumulation of tracers. SUV can be normalized to body mass, lean body mass (SUL), or body surface area. Body surface area and SUL are less dependent on body habitus across populations than is SUV based on total body mass. In a single patient of stable weight, all 3 SUV normalization approaches will give comparable percentage changes with treatment, as the normalization terms cancel out mathematically. However, the absolute change in SUV with effective treatment and the absolute amount of change in SUV to be significantly different from a prior scan will differ on the basis of the metric used.

The determination of SUV is dependent on identical patient preparation and adequate scan quality that is similar between the baseline and follow-up studies. Ideally, the scans should be performed on the same scanner with comparable injected doses of ¹⁸F-FDG and comparable uptake times before scanning. Absolute and rigorous standardization of the protocol for PET is required to achieve reproducible SUVs. Standardization has been well summarized in a consensus document from the National Institutes of Health and a recent report from The Netherlands (30,31). SUL is preferred by many over SUV normalized by body surface area, as the SUL values are relatively close to (though usually somewhat less than) SUVs normalized on the basis of total body mass (30,102,103). SUL is typically more consistent from patient to patient than is total-body-mass SUV, as patients with high body mass indices have high normal organ SUVs because ¹⁸F-FDG does not significantly accumulate in white fat in the fasting state (102,103).

ROI selection is a key aspect of determining tumor SUV, tumor Ki, or any quantitative PET parameter. A wide variety of SUV ROI selection metrics has been used: manually defined ROIs; irregular isocontour ROIs based on a fixed percentage of the maximal pixel in the tumor (e.g., 41%, 50%, 70%, 75%, or 90% of the maximum); irregular isocontour ROIs based on a fixed SUV threshold (e.g., SUV = 2.5); irregular isocontour ROIs based on a background-level threshold (e.g., relevant background + 2–3 SDs); and small fixed-dimension ROIs centered over the highest-uptake part of the tumor (e.g., 15-mm-diameter circles or spheres or 12 × 12 mm squares, giving rise to a parameter sometimes called SUV peak). In addition, SUV is frequently obtained from the pixel with the SUVmax and, although not usually determined in this way, it could be considered to be a single-pixel ROI.

As part of this special contribution, we have ascertained the methods for ROI selection in determining SUV in cancer studies in over 1,000 reports. The use of varying regions of interest to determine SUV over the past decade is shown in Figure 2. It is apparent that SUVmax is growing in use and is the de facto standard, given its widespread use. A close examination of the graph shows a growing use of SUV peak, as well. The isocontour and manual ROIs have also been applied in some studies. Given that the use of SUVmax is so commonly reported, it might seem to be the “best” method. However, the wide use of SUVmax may also be due its being easily measured using current commercial workstations. To simply recommend SUVmax as the preferred treatment response parameter would be easy, as it should also be most resistant to partial-volume issues in small tumors. However, this recommendation must be taken with some trepidation as SUVmax is highly dependent on the statistical quality of the images and the size of the maximal pixel (104). For SUVmax to be used routinely, its performance characteristics should be well understood, including its reproducibility versus other approaches.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2. 

Number of papers that included use of tumor ROIs, as function of year of publication. Papers were identified by Medline search that queried for FDG AND SUV OR “standard uptake value” OR “standardized uptake value” OR “standardised uptake value”). Only human ¹⁸F-FDG oncology studies were included. ROI max refers to maximal pixel in tumor. ROI peak refers to small (typically 15 × 15 mm) fixed-size ROI centered on most metabolically active part of tumor. ROI isocontour refers to irregular ROI defined by isocontour set at, for example, some percentage of maximal pixel. ROI manual refers to manually drawn ROI. Only a subset of these papers describes response assessment studies.

A fundamental biologic question underlying choices of regions of interest is whether the total tumor volume or the maximally metabolically active portion of the tumor is most important. Intuitively, both would seem important and desirable to determine. However, concepts of stem cell biology suggest that the most critically important parts of tumors are the most aggressive portions, which may not be the entire tumor. This controversial concept is under study for many cancers (105–108). In practice, much of the early development of PET for treatment response was in the setting of a single tumor, as neoadjuvant therapy or as palliative treatment. Most papers focus on a single or a few tumor foci in ROI selection. However, the total lesion volume and its metabolic activity, known as the total lesion glycolysis, effective glycolytic volume, or total glycolytic volume (calculated in similar manners—mean SUV of the total tumor times × total tumor volume, in mL), are potentially important parameters for studying the behavior of the total tumor (109–112). For the purposes of this article, although the terms represent similar indices, we will refer to total lesion glycolysis in discussions of response based on total lesion volume and its metabolic activity.

To use quantitative metrics to assess treatment response, one must know their performance characteristics. We are aware of 5 reports on the test–retest reproducibility of PET with ¹⁸F-FDG in cancer, and the major methods and protocols of these studies are summarized in Table 6 (100,113–115). Overall, the reproducibility of quantitative PET parameters in the test–retest setting has varied depending on lesion size and the methods for image acquisition, reconstruction, and analysis. The lowest variability in PET quantitative parameters is in the 6%−10% range, but up to 42% variability has been reported. In the test–retest setting, ROI and lesion size seem to be important for SUV reproducibility whereas reproducibility appears less dependent on glucose correction factors (113,114) and the reconstruction method used (filtered backprojection vs. ordered-subset expectation maximization) (100).

View this table:

• View inline
• View popup

TABLE 6

Summary of Studies on Test–Retest Reproducibility of Untreated Tumors Without Interval Therapy

Minn et al. (116) first demonstrated that although kinetic modeling with nonlinear regression is conceptually more attractive than SUV, it is not as reproducible in the test–retest setting as is the simpler Patlak-derived Ki or the SUV. Because both Ki and SUV (or SUL or body-surface-area SUV) correlate well with kinetic modeling results, full kinetic modeling approaches are not typically undertaken in treatment response monitoring with ¹⁸F-FDG.

Ki is an attractive parameter and may be helpful when the SUV after treatment is low (117). However, Ki requires a period of dynamic scanning, a process typically more time consuming and restricted in the spatial location evaluated than whole-body PET. Further, only limited standard software is available for generation of Ki values.

The size of the ROI affects the reproducibility of SUV. SUVs obtained from larger, fixed ROIs are more reproducible than single-pixel SUVs (110,115, 118). Comparing the test–retest studies in Table 6, one can see that the ROI used by Minn in 1995 (113) was 39-fold larger in volume than that used by Nahmias and Wahl (115) in 2008 for single-voxel SUVmax (438 mm³ vs. 12.5 mm³). For equal sensitivity, there would be 39-fold fewer counts in the maximal pixel using modern PET scanners, versus the volume applied originally in determining the statistical precision of PET in the test–retest setting using older equipment with thicker slices and smaller matrices.

The assessment of Nakamoto et al. (110) of the data of Minn et al. (113) used a smaller maximal pixel volume, but it was still about 19 times larger than the volume of a single voxel used in many current scanners. Weber et al. (114) used regions of interest much larger than those of Minn et al., presumably increasing statistical reliability. Further, data from Nahmias and Wahl (115) were obtained at 90 min after injection and not the 50- to 60-min time used by Minn (113), meaning radioactive decay further reduced the total counts.

Reproducibility data from individual patients are likely of greatest practical interest in evaluating the degree of change required to determine that a change is significant between 2 studies. Weber et al. (114), using a larger ROI, reported that 0.9 SUV unit was needed for a significant change. Concordantly, Nahmias and Wahl (115) showed in test–retest studies that absolute differences in mean SUV obtained from a large ROI did not exceed 0.5 SUV unit and that the absolute differences in mean SUV decreased as mean SUV increased. In contrast, the absolute difference between SUVmax increased to over 1.5 SUV units in a substantial number of cases in which the SUVmax was over 7.5 (i.e., the hotter tumors). Thus, there are differences in the behaviors of SUVmax and mean SUV in terms of reproducibility that likely will have a direct impact on the fractional and absolute changes required to have a significant difference between a baseline and a follow-up scan.

The large ROI of Nahmias (115) showed superb test–retest performance; however, the size of their circular ROI was both manually determined and manually positioned, and thus it may be difficult to routinely achieve such low variability at other centers. Larger ROIs may be too big for small tumors such as nodes to be optimally assessed, as well.

These human data are augmented by phantom and modeling data. Boellaard et al. also showed that SUVmax variability increases as the lesion matrix size is increased from 128 × 128 to 256 × 256. They also showed that the variability increases with lower counts as the patient size increased (and the statistical quality decreased) (104).

The appeal of the single maximal pixel value is undeniable, but it is clear that with modern scanners and many small voxels, it is not as reproducible as larger ROIs and that larger changes in SUVmax between studies are needed for significance (104). This is mainly because of noise effects on SUV, which induce a positive bias in the recovery coefficient for SUVmax. As lesions get larger and hotter, there is also a statistical bias to higher single-pixel SUVmax simply because of the number of counts available. This raises concern, especially given the widespread and growing use of this parameter in clinical studies with PET, and caution must be applied in the use of single-pixel SUVmax for assessing small changes induced by treatment. For these reasons, it is probably important to have a minimum ROI for PET metrics of maximal tumor activity to ensure adequate statistical quality and intrastudy comparability.

Methods for determining total lesion glycolysis are still evolving. Choosing a threshold based on a single maximal pixel value in the tumor carries with it the variability inherent in determining a single-pixel value and is driven by that value (104,109,112,119,120). Investigators have also found poor reproducibility for tumor volume estimates (also applied to calculate total lesion glycolysis) using thresholding methods based on the maximal pixel value. After treatment, thresholding methods for tumor volume determination may extend to include too much normal tissue (118). The use of thresholds such as “anything 3 SDs or greater above background is tumor” is one approach that has been applied to defining lung cancer volumes on PET, avoiding the uncertainty of SUVmax (121). A background threshold approach has been developed as a tool for defining metabolic tumor volumes for mesotheliomas with good initial success, choosing 3 SDs above background levels for segmentation (111). Other approaches include determining the lesion volume not from PET but from the CT of the PET/CT (122). These methods hold great promise for providing the tumor burden, which may be quite important as a complement and addition to SUV.

One other approach, akin to total lesion glycolysis, is the multiplication of SUVmax × tumor width to provide a combined glycolysis × size parameter. Such approaches may be useful in response assessment but have not been extensively assessed. They could suffer from the variance intrinsic in the metabolic and anatomic methods, potentially reducing the precision of the methods, but initial results are encouraging in esophageal cancer treatment assessment (123).

Comparing tumor activity to background is an attractive way to minimize variability and to potentially ensure the quality of scans from test to retest. A variety of backgrounds has been used. Thighs, back muscle, liver, and mediastinum, for example, have been measured. Pacquet et al. showed that liver SUV is quite stable over time, when measured as a mean on a single slice in the right lobe of the liver centrally, as is mean mediastinal blood pool (124). Paquet et al. reported that mean SUL in the mediastinum was 1.33 ± 0.21 and 1.30 ± 0.21 (within-patient coefficient of variation, 12.3%) on test–retest. Mean SUL in the liver showed slightly less variance (within-patient coefficient of variation, 10.8%) and was 1.49 ± 0.25 and 1.45 ± 0.20. Glucose correction and use of the SUVmax in the liver or blood pool resulted in considerably higher variance and were not recommended for normalization. Similar results for normal organ uptakes were reported by Minn et al. in limited tissues, as well as by Wahl et al., among others (20,113). These values were slightly higher than mean blood-pool values. Krak et al. recommended the use of SUL for monitoring treatment response, as well, although they favored glucose correction (100).

A variety of methods has been used to determine the change in SUV with treatment. SUVmax in a single pixel, background-corrected values, larger or smaller ROIs, and total lesion glycolysis have been used, among others. The prospective data of Weber et al. are among the most compelling (125). Based on the differences seen in test–retest studies, they evaluated changes in SUV in cases that met the following characteristics: tumor clearly visible, large enough, and hot enough (2 × blood-pool background). Using a 1.5-cm ROI, they showed in lung, gastric, and esophageal cancers that declines in ¹⁸F-FDG uptake of 20%−35% after 1–2 doses of therapy are predictive of outcomes, with the larger the drop, the greater being the beneficial effect. In esophageal cancer, for example, Weber et al. found a drop of greater than 35% in SUV to be a good predictor of response (125). In neoadjuvant gastric cancer therapy, in which tumors with an SUV of more than 1.35 times the mean liver SUV + 2 SDs were assessed, the mean decline in SUV was about 50% in responders and 18% in nonresponders (126).

Weber has argued that any drop of more than 20% is significant and should be called a response on the basis of reproducibility considerations (Radiological Society of North America syllabus). However, in most studies, larger drops in SUV of more than 30%−35% are seen and associated with a good outcome. In lymphomas, at mid therapy, a drop in SUV of 65.7% was best at separating favorable from unfavorable responses and appeared superior to visual examination (accuracy visual, 65.2%; SUV reduction, 76%; tumor-to-background ratio, 74%; and SUV floor, 74%) in a study by Lin et al. (86). Although quantitative analysis appeared superior to visual analyses (though it must be cautioned that this was using a retrospective cutoff value and there was considerable overlap in the best responding and less well responding groups quantitatively—as well as a fine continuous scale for quantitation but a coarser approach for visual analyses), the several quantitative approaches appeared quite comparable. The authors favored the percentage decline in SUV. It appears that many methods of quantification can produce valuable prognostic information on treatment response using PET.

Another issue in PET treatment response is whether an absolute SUV floor or threshold (such as blood-pool background in the non-Hodgkin lymphoma PET criteria) or a percentage decline in SUV is most important. The advantages to a percentage drop in SUV versus a floor are that the percentage drop is likely easier to calculate than the absolute SUV; many measurement issues become less important when test–retest studies are done, because the technical issues are constant across studies. Modeling studies have shown that the ratios of SUV are less dependent on ROI choice than are absolute SUV determinations (104). An SUV floor carries the advantage of allowing a baseline PET scan to be obtained at another center to verify the ¹⁸F-FDG avidity of the tumor, but such a baseline study is not required for quantitation.

The data of Lin et al. (86) show nearly comparable results for floor SUV versus percentage decline in terms of ability to separate those with a good response from those with a less good response to treatment for non-Hodgkin lymphoma. However, several papers have shown that in lung cancer, for example, a decline in a tumor SUV to below 4–6 after treatment separates groups of patients with longer and shorter survival reasonably well (72,127). The differing cutoffs suggest possible differences in SUV calculation approaches. Reproducing absolute SUV across centers can be difficult, however, and although such absolute cutoffs may be valuable for determining prognosis, they are viewed as more suitable in single-center studies or in well-controlled multicenter approaches using careful standardization methods (31). It may be possible to determine a simple floor for PET through the use of normalization to structures such as the normal liver or blood pool, for example, as has been done qualitatively in the IWC + PET criteria (33).

SUVs in normal tissues are not stable with time, because blood-pool and liver uptake fall with increasing delays from injection, whereas uptake in tumor typically rises (20,128). Thus, normalization is difficult if scan uptake times vary. However, a threshold for posttreatment PET is an attractive concept and may be more important in the future as standardization for PET performance improves.

Methods of assessing response to treatment with total lesion glycolysis are still evolving. It appears that percentage declines in total lesion glycolysis are sometimes greater than declines in SUV and that total lesion glycolysis gives a larger range of changes after treatment than does SUV (111). This would suggest that larger changes in total lesion glycolysis would be required to have a meaningful response than are required for SUV alone. Francis has found total lesion glycolysis to be superior to SUVmax in mesothelioma response assessment. However, SUVmax is also a potent predictor of outcomes in other studies of mesothelioma (52,129) and is quite strong in the data of Francis et al., as well (111). In studies of colorectal cancer neoadjuvant response, SUVmax appeared to perform somewhat better than total lesion glycolysis, though it depended on the specific task involved (57). Total lesion glycolysis has performed well in studies of colorectal cancer and brain tumor response (109,112,119,120). In studies of sarcoma response, total lesion glycolysis performed less well than SUV peak (122). Thus, the total lesion glycolysis parameter appears promising in some, though not all, cancers. The method by which it is calculated can be quite variable, however.

The EORTC PET response criteria were proposed in 1999 (36). Given the limited data available on treatment response at that time, the criteria were useful and prescient. They recognized that the subclinical metabolic response seen early after treatment on PET, but not seen anatomically, was likely to be important. The group made several important points in its report regarding the ¹⁸F-FDG PET response: careful methods and patient preparation are essential; early declines in SUV with effective therapy will be smaller than later ones; with ineffective treatment, tumors can progress not only by increasing their SUV but also by physically growing; accurate and reproducible methods are essential for accurate reporting; and as the literature matures, updates will be needed (36).

Drawing from their work and the maturing literature on treatment response assessment over the intervening decade, some additional suggestions regarding treatment response criteria are in order.

Introduction to PERCIST 1.0

Based on the extensive literature now supporting the use of ¹⁸F-FDG PET to assess early treatment response as well as the known limitations of anatomic imaging, updated draft PET criteria are proposed that may be useful for consideration in clinical trials and possibly clinical practice. We have called these draft criteria “PERCIST”—Positron Emission tomography Response Criteria In Solid Tumors. The RECIST committee did not have a role in developing these criteria, but while we were developing them we acknowledged and appreciated the careful work and approaches of the RECIST committee. We also recognized that, as with RECIST, criteria such as PERCIST will need updates and validation in differing settings. With apologies to the RECIST group, we believed that the name PERCIST seemed quite appropriate as a complement to the well-developed anatomic criteria now in widespread use and recently updated.

The premise of the PERCIST 1.0 criteria is that cancer response as assessed by PET is a continuous and time-dependent variable. A tumor may be evaluated at any number of times during treatment, and glucose use may rise or fall from baseline values. SUV will likely vary for the same tumor and the same treatment at different times. For example, tracer uptake by a tumor is expected to decline over time with effective treatment. Thus, capturing and reporting the fractional change in SUV from the starting value and when the scan was obtained are important.

The optimal number of chemotherapy cycles before obtaining an ¹⁸F-FDG PET scan and the optimal interval between the last treatment and the scan are matters of debate and may be treatment-specific. Our assessment of the literature and the conceptual framework in Figure 1 suggest that early after treatment (i.e., after 1 cycle, just before the next cycle) may be a reasonable time for monitoring response, to determine whether the tumor shows no primary resistance to the treatment. Indeed, several studies, including one by Avril et al. on ovarian cancer, show that 60%−70% of the total SUV decline occurs after just 1 cycle of effective treatment (130). By contrast, waiting until the end of treatment can provide evidence that resistance to treatment was present throughout the treatment or evolved during treatment. End-of-therapy PET scans are quite commonly performed as restaging examinations to determine whether additional treatment or possibly surgery should be performed.

After chemotherapy, waiting a minimum of 10 d before performing ¹⁸F-FDG PET is advised. This time permits bypassing of the chemotherapeutic effect and of transient fluctuations in ¹⁸F-FDG uptake that may occur early after treatment—stunning or flare of tumor uptake (131–133). The guidelines of the IWC + PET criteria for lymphoma recommend waiting at least 3 wk between the last chemotherapy session and ¹⁸F-FDG PET, but we recognize that this longer waiting period might not be feasible for all cases. Longer and more variable times after external-beam radiation, 8–12 wk, have been recommended (134).

The basics of PERCIST 1.0 are shown in Table 7, where they are contrasted with the EORTC criteria. Key elements of PERCIST include performance of PET scans in a method consistent with the National Cancer Institute recommendations and those of The Netherlands multicenter trial group (30) on well-calibrated and well-maintained scanners. Patients should have been fasting for at least 4–6 h before undergoing scanning, and the measured serum glucose level (no correction) must be less than 200 mg/dL. The patients may be on oral hypoglycemics but not on insulin. A baseline PET scan should be obtained at 50–70 min after tracer injection. The follow-up scan should be obtained within 15 min (but always 50 min or later) of the baseline scan. All scans should be performed on the same PET scanner with the same injected dose ± 20% of radioactivity. Appropriate attenuation correction along with evaluation for proper PET and CT registration of the quantitated areas should be performed.

SUV should be corrected for lean body mass (SUL) and should not be corrected for serum glucose levels (glucose corrections have been variably useful, and errors in glucometer measurements are well known and may add errors (135)). Normal background ¹⁸F-FDG activity is determined in the right hepatic lobe and consists of mean SUL and SD in a 3-cm-diameter spherical ROI. Typically, liver uptake should not vary by more than 0.3 SUL unit from study to study.

The SUL is determined for up to 5 tumors (up to 2 per organ) with the most intense ¹⁸F-FDG uptake. These will typically be the lesions identified on RECIST 1.1. The SUV peak (this is a sphere with a diameter of approximately 1.2 cm—to produce a 1-cm³-volume spheric ROI) centering around the hottest point in the tumor foci should be determined, and the image planes and coordinates should be noted (SUL peak). This SUL peak ROI will typically include the maximal SUL pixel (which should also be recorded) but is not necessarily centered on the maximal SUL pixel. Automated methods for searching for this peak region have been described (20). Tumor sizes should be noted and should be 2 cm or larger in diameter for accurate measurement, though smaller lesions of sufficient ¹⁸F-FDG uptake, including those not well seen anatomically, can be assessed. Each baseline (pretreatment) tumor SUL peak must be 1.5 × mean liver SUL + 2 SDs of mean SUL. If the liver is diseased, 2.0 × blood-pool ¹⁸F-FDG activity + 2 SDs in the mediastinum is suggested as minimal metabolically measurable tumor activity.

In PERCIST, response to therapy is assessed as a continuous variable and expressed as percentage change in SUL peak (or sum of lesion SULs) between the pre- and posttreatment scans. Briefly, a complete metabolic response is defined as visual disappearance of all metabolically active tumor. A partial response is considered more than a 30% and a 0.8-unit decline in SUL peak between the most intense lesion before treatment and the most intense lesion after treatment, although not necessarily the same lesion. More than a 30% and 0.8-unit increase in SUL peak or new lesions, if confirmed, is classified as progressive disease. A greater than 75% increase in total lesion glycolysis is proposed as another metric of progression. Further details of the proposed PERCIST criteria for monitoring therapy response and comparison to EORTC are shown in Table 7.

Why PERCIST?

PET assessments of treatment response with ¹⁸F-FDG appear to have substantial biologic relevance when obtained at the end of treatment, at mid treatment, or soon after treatment is started. Indeed, the biologic predictive value of PET appears to be greater than that of anatomic studies, including for lymphoma, lung cancer, mesothelioma, and esophageal cancer. Although currently accepted response criteria are anatomic, it is quite possible that an approach using purely metabolic response criteria may ultimately be more predictive of outcomes. Given that some tumors do not have high uptake of ¹⁸F-FDG, or may be too small to be reliably quantified, it is likely that both anatomic and functional criteria will be important for the foreseeable future. Although it would be possible to propose an integrated CT + PET approach akin to that of the IWC + PET (i.e., that a PET scan only be interpreted as positive or negative and be used to trump anatomic imaging if the studies are disparate), this approach would seem to lose some of the advantages of the continuous output of the PET data through forced dichotomization. The inclusion of an ¹⁸F-FDG PET observation into the RECIST 1.1 criteria as a sign of disease recurrence is a step in this direction.

In preparing the PERCIST 1.0 criteria, at the request of The Journal of Nuclear Medicine editors (after the lead author had lectured on this topic), it was clear that many of the answers regarding the use of PET for assessing treatment response are not yet in. What is clear is that unless more precisely defined response criteria are in place and used by varying groups, it will be difficult to compare PET treatment response studies across centers or even to include PET in such studies. The Imaging Response Assessment Team at Johns Hopkins reviews clinical oncologic protocols at the Sidney Kimmel Comprehensive Cancer Center weekly. In nearly all of these, RECIST criteria are used for solid tumor evaluations. Only a few studies include PET. Although some use the EORTC criteria, methods for PET performance and interpretation are typically highly variable across studies and typically only exploratory. With over 30 ways to assess tumor response quantitatively and many articles using differing ROI selection techniques, arriving at a common approach, even if not proven ultimately to be the best in each case, will help generate more data on treatment response and allow a larger database to be developed for testing analytic tools retrospectively as has been done by the RECIST group.

Why the ROI?

Several points in the PERCIST 1.0 criteria are notable and may be controversial. ROI size is important and has varied from study to study. Larger ROIs give better precision but a lower SUL than do smaller ROIs (20,115). Despite its widespread use, maximal SUL was not selected as the primary metric of response because the size of the maximal voxel sampling ROI varies considerably by scanner, matrix size, slice thickness, and scanner diameter, resulting in various noise levels in the metric. Thus, the precision of maximal SUL is not well established. All but one of the studies examining the precision of SUV used larger regions of interest than the volume assessed to determine the current single-pixel SUVmax provided by modern high-resolution scanners. When tested, the small single-pixel SUVmax is more variable than the somewhat larger ROIs.

The maximal pixel value is possibly most advantageous in small tumors, as it would be somewhat less dependent on partial-volume effects. However, noise effects are substantial. Although correcting for partial volume is attractive conceptually, the PERCIST criteria have avoided partial-volume corrections. Measuring tumor or node size with CT from PET/CT is feasible, but slight errors in those measurements can have major effects on quantitation if used to correct for partial-volume effects. Studies in which complex partial-volume corrections have been performed in addition to corrections for background spillover from nearby tissues have sometimes, but not consistently, demonstrated quantitation to be superior to visual assessments for predicting response and outcome (136). We believe such corrections will be to too difficult to effect in routine practice because of the obvious challenges of measuring small lesions accurately. The maximal SUL should be recorded, however, for selected ¹⁸F-FDG–avid tumors.

Most studies of treatment response have focused on larger measurable tumors. We realize maximal SUL may be useful in small lesions and should be explored. Although imaging tumors larger than 2 cm is encouraged to minimize partial-volume effects, PERCIST 1.0 allows any tumor whose SUL peak is greater than 1.5 × liver mean + 2 SDs to be assessed quantitatively. This figure is based on cutoffs used by Weber and is used to ensure that the posttreatment lesion SUL can fall sufficiently to detect a response. Less avid tumors may be visualized and their disappearance can be noted, as well as their obvious progression. It is possible that a cutoff of 1.35 × hepatic uptake as was used by Weber may also be acceptable as a lower limit of measurable activity.

However, recording tumor size by RECIST criteria is suggested for measurable lesions larger than 1 cm. Because ROIs whose size is based on a 50%, or other, ROI threshold vary with the variability of the maximal pixel chosen, these were not chosen as the primary measurement metric. Rather, the SUL peak in a small volume of greatest metabolic activity in the tumor (approximately 1 cm³) is suggested for use. This size has been used in many studies and is statistically less subject to variance than is a small, single-pixel SUVmax.

Total lesion glycolysis is also attractive. PERCIST suggests that this be obtained but recommends that it be threshold-based, with an outer boundary equal to 3 SDs above normal-liver mean SUL determined in a standard-sized ROI of 3 cm in diameter. This should be relatively consistent, based on such factors as similar injection times for imaging on the baseline study and the follow-up study. However, the total lesion glycolysis metric is not proposed for primary response assessment. We suggested that it be routinely obtained for the 5 hottest lesions to estimate tumor burden, but it is optional for assessing all lesions. Collecting these data consistently should help us learn more about the best method to assess treatment response by disease type.

What Decline in SUV Is a Response?

Already, it is evident that the medically relevant cutoff for an SUL decline to represent response and predict outcomes may differ on the basis of the disease, the timing after treatment, the treatment itself, and the treatment goal. The 30% requirement for a tumor response (and the drop of 0.8 SUL unit) we propose in PERCIST (based on peak SUL) is more stringent than that proposed in the 1999 EORTC criteria (15% or 25% drops in SUV). The 15% decline in SUV in the original EORTC criteria for early response is probably too modest to reliably be discerned from variability in the study and likely is insufficient to be medically relevant based on data developed since that time.

For lymphomas, in which cure is feasible and a rapid drop in SUV is common, a higher cutoff for a medically relevant response (e.g., 65% at mid treatment) may be required (86). This cutoff is greater than that for the palliative or noncurative treatment of lung cancer (e.g., 30%−35%). Similarly, in sarcoma and gastric and ovarian carcinoma responses, a drop in SUV of more than 25% is associated with the best outcomes (43,87,137,138). When lower thresholds of, for example, 20%−30% are accepted as responses, limited data suggest that these patients are unlikely to have a medically relevant response, even if the response is statistically significant (87,130). For example, patients with GIST treated with imatinib who had only modest declines (∼30% decrease) in SUV early after therapy did not appear to have good outcomes, suggesting that a larger threshold may have been in order (87).

Although a decline of 25% or more is less likely to be due to chance than are smaller declines, this level of decline can occur in lesions with low SUVs and a rather modest change in total SUV. For this reason, a minimal level of tumor uptake is proposed in PERCIST 1.0 to be assessable. This minimal level is proposed as 1.5 × the liver SUV mean + 2 SDs. Because the typical SUL of the liver is around 1.6–1.8, the SUL peak of an assessable lesion is going to be approximately 2.5 or greater (Fig. 3). In addition to the requisite percentage change in SUL after treatment, PERCIST also requires a defined absolute change in SUL of 0.8 units in order to minimize overestimation of response or progression. Weber has proposed a 0.9 SUV change as the minimum to be significant (114); however, since SUL is typically somewhat less than SUV, we suggest a change of 0.8 SUL unit to be a reasonable absolute change. The 0.5 SUV unit change described as significant by Nahmias (115) may be too small with the ROI size proposed for PERCIST. We do not know what change in total lesion glycolysis is required for a response. Because the dynamic range is larger, a suggested figure of 40% for a response should be considered on the basis of the larger changes in total lesion glycolysis than SUVmax reported in mesothelioma, as well as a potentially, but not fully defined, lower precision for the volume × SUV figure, which would be expected because of measurement errors in both the volume and the SUV parameters (111).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3. 

Example calculation of liver background for normalization of SUL. Images are displayed from Advantage Workstation (GE Healthcare). A 3-cm-diameter 3-dimensional ROI (ROI 1) is placed on normal inferior right lobe of liver (arrowhead). Average SUL and SD in ROI are displayed (arrows). Liver background is calculated as follows: (1.5 × average SUL liver) + (2 × SD average SUL liver). For this example, (1.5 × 1.4) + (2 × 0.2) = 2.5. Therefore, tumor SUL peak should be >2.5 in order to apply PERCIST criteria for this example.

It is also important in PERCIST to note how long into the therapy the response is obtained to take full advantage of the continuous nature of the SUV. Recording of the full continuous range of the percentage change in SUL allows for preservation of data that are otherwise lost by reducing the continuous variable to discrete bins of response.

Using continuous data, it should be possible to perform controlled trials in which experimental treatments are compared with standard treatments. In such trials, the expected change in SUL may not be known. However, the continuous readout of SUL change is expected to be quite helpful in detecting the activity of the therapeutic agent and to minimize sample sizes.

The PERCIST 1.0 criteria are designed to facilitate trials of drug development but, if sufficiently robust, could be applied to individual patients. In individual patients, determining what level of quantitative change in SUL is medically significant will depend on multiple factors, not just on what level of change exceeds that due to chance. Other factors will include the level of comfort the treating physician has in not treating with a regimen that may still have a small likelihood of being effective (i.e., of deciding to deny therapy to someone who may have a borderline response and a low, but possible, chance of benefit). Decisions to deny probably ineffective therapy depend on alternative therapeutic options and on the risks, cost, and benefits of the treatment and so are difficult to specifically address. If therapies are of low risk and there are no good alternatives, denial of treatment would seem unreasonable, even if benefit were quite improbable. By contrast, with a highly toxic treatment of high cost, denying treatment might be highly appropriate if the treatment is unlikely to be beneficial. As more data are generated on specific diseases with specific treatments, the development of likelihood ratios of probable benefit from treatment can be expected. An example of a partial metabolic response by PERCIST is shown in Figure 4, one in which the functional response exceeds the anatomic.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4. 

PET/CT images obtained before (1) and after (2) treatment of pancreatic carcinoma with experimental therapy targeting mammalian target of rapamycin. Note profound decline in SUL (∼41%) despite stable pancreatic mass anatomically (arrows). This decline represents metabolic partial response by PERCIST (41% decline in marker lesion at 2 wk after therapy). Not all metabolic PMRs are clinically relevant; relevance will depend on the specific treatment.

What Decline in SUV Represents a Complete Response?

The PERCIST criteria do include the category “complete metabolic response.” It might seem logical that patients with a complete response would have a 100% SUV decline. However, in many studies the degree of SUV reduction associated with a complete metabolic response is less than 100% (139). PERCIST specifies that the SUL percentage reduction be noted from the pretreatment to the posttreatment PET scans, along with the time from the start of the most recent treatment regimen (in weeks), even for complete response in patients on active treatment. Because background rarely has an SUL of 0, declines in SUL to 0 are unlikely, as are 100% reductions in tumor SUL.

Drops in SUL of 100% could be achieved by subtracting the mean SUL of the liver + 2 SDs from the tumor activity and using the resultant dynamic range. However, after treatment, drops in SUL of over 100% are possible with such an approach. For small lesions after treatment, focal uptake may remain and may be less than liver uptake and visually detectable (32). Thus, the possibility of a incomplete response with over a 100% decline in background-corrected SUL exists. PERCIST 1.0 requires collection of the background SUL in the liver and the variance in SUL, which can allow for such post hoc calculations of background-corrected SUL changes if desired. For this PERCIST 1.0 version, we believe visual assessment is essential for determining the presence or absence of complete response, especially for small lesions after treatment. However, data collected from our approach should allow future studies of the best definition of complete response to help define whether a qualitative or quantitative metric is most robust at predicting outcomes. Quantitative metrics potentially may be developed to help in avoiding false-positive scans after treatment.

What About the Choice of Background?

Background tissues are important normal metrics for verifying that a PET study is performed properly from a technical standpoint. Many factors, including a poor intravenous injection, inaccurate dose calibration or camera calibration, or variable uptake times, can affect the SUL (30). We believe that the normal liver SUL is slightly more stable than determinations of blood-pool SUL. Practically, it is less effort to draw a 3-cm-diameter ROI on the right lobe of the liver than to repeatedly draw regions of interest on the aorta on multiple levels, taking care to avoid including uptake in the possibly diseased vessel wall (113,114,124,140). If the liver is diseased (most notably, full of cancer involvement), it is clearly unsuitable as a background area. An alternative in such a case is the blood-pool activity in the descending aorta. For either blood pool or liver, the SUL temporally depends on the time after injection. Thus, close similarity in uptake times is required for the baseline and follow-up studies to ensure the stability of background hepatic uptake.

How Many Lesions to Assess?

The number of lesions to evaluate when assessing response to therapy is a major issue, and the answer is uncertain for PET at this time. Most of the initial PET literature evaluated a single lesion, such as a primary lung, breast, or esophageal cancer. In such cases, n = 1 is obviously the appropriate number. In anatomic imaging assessments in which multiple tumors are present, the RECIST group has recently recommended evaluating the size of a maximum of 3–5 lesions (typically 5) anatomically to assess response, even if many more lesions are present. This does not mean other lesions are not assessed; rather, it means they are not measured. If tumors other than these 5 progress unequivocally, progression has occurred (39,40). RECIST separates between target and nontarget lesions (Tables 1 and 2).

In the Hicks qualitative PET criteria (Table 5), multiple lesions are assessed (76,84,92,141). In quantitatively assessing treatment response in patients with disseminated ovarian cancer, Avril et al. assessed up to 4 lesions per patient, but an average of just 2.2 lesions were studied for response (130). They chose the lesion with the smallest percentage decline in SUV after therapy as representative (i.e., the worst responder), with a rationale that the metastatic tumor with the worst response would determine survival.

In another study of disseminated intraabdominal tumors, Stroobants et al. selected up to 3 foci of ¹⁸F-FDG uptake in GIST that were highest on baseline PET. All lesions had to decline by at least 25% to represent a partial response, and all had to disappear to background to represent a complete response (87).

Remarkably, several studies have shown that changes in the SUV of primary tumors can quite accurately predict the outcomes in their nodal metastases. Careful studies from Dooms et al. have shown that metastatic-tumor-involved mediastinal nodal pathology and clinical behavior are well predicted by changes in SUV and absolute SUV in the primary lung cancer and by qualitative visual assessments of nodal status (66,142). This is in part because “child” metastases biologically resemble their “parents” (143,144).

Several other interesting approaches have evaluated just a single lesion but considered the worst-case biologic behavior of the malignancy. Lin et al. found that the accuracy of predicting event-free survival in lymphoma response assessment was slightly better using the change in SUV from the hottest lesion on study 1 to the hottest lesion on study 2 (which was a different lesion in 18% of cases) than using the change in the hottest lesion on the baseline study (76.1% accuracy vs. 73.9% accuracy in outcome prediction) (86). Although comparable, there were slightly more false-negative scans when the same lesion was used for analysis. This approach is somewhat similar to that used by Wahl et al., in which the single hottest area in a primary breast cancer was used as the reference point on the pretreatment and posttreatment studies—often, but not necessarily, the same area (20).

Because the RECIST criteria examine a maximum of 5 lesions, we have proposed that PERCIST measure the SUL in no more than 5 lesions, as well (unless an automated total lesion glycolysis is determined as a corollary study). However, it is not known how to optimally combine the results of percentage change in SUL from multiple tumors to be predictive of outcome. For example, to have a response, does each metabolically assessable target tumor have to drop its uptake by 30%, or does the sum of the declines in SUL in the posttreatment group have to be 30% less than the sum of the SULs in the same lesions before treatment? Requiring each lesion to drop at least 30% is probably more stringent than the sum, but this is not clear. It is probable that combination methods of either summed SUL before and after treatment (sum of SUL for lesions 1–5 before treatment and sum of SUL of lesions 1–5 after treatment) or percentage decline in summed SUL between scans will be biased by the hottest lesion or largest percentage decline.

The uncertainty on how to precisely combine the SULs of 5 lesions, and evidence that a restricted dataset of fewer tumors is commonly adequate, along with simplicity of calculation are other reasons why, for this first-level analysis of PERCIST 1.0, it is suggested that only the percentage difference in SUL between the tumor with the most intense SUL on study 1 and the tumor with the most intense SUL on study 2 should be used as a classifier for response. This suggestion supposes that the most intense lesion on study 2 has not grossly progressed and that it was present at the time of study 1. As long as all other unmeasured lesions do not progress, this method would be used to determine whether a response had occurred. Given the uncertainty about the best metric, it is suggested that SUL peak data be determined and summed before and after treatment for up to the 5 hottest lesions and that the ratio of the sums before and after treatment be compared as a secondary analysis. Obvious progression of any tumor (i.e., >30% increase) or new lesions would negate a partial response.

Perhaps these findings that one or a few tumors predict outcome well are consistent with the clonality of metastases; that is, most are genetically comparable and most respond similarly to treatments. Thus, a good assessment of the most metabolically aggressive tumor before and after treatment may be reflective of the others in many instances. However, we all have observed cases in which new lesions appear and progress despite apparent control of a primary lesion (Fig. 5) (139). This observation may be related to the form of treatment but does occur. Thus, clearly progressive disease in any one lesion is disease progression, even if other tumor foci are responding.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5. 

PET/CT image obtained before (1) and after (2) treatment of pancreatic carcinoma with experimental therapy targeting mammalian target of rapamycin. Glycolysis and apparent necrosis are profoundly reduced in intensely ¹⁸F-FDG–avid liver metastases. Although a reduction of more than 50% in SUL peak would suggest partial metabolic response, new lesion indicative of progressive metabolic disease is evident in left retroperitoneum (arrow).

Lack of Good Information for Progression

The precise optimal definition of tumor progression remains in evolution. The EORTC criteria defined progression as an increase in SUV of over 25%, an increase in the extent of ¹⁸F-FDG uptake by more than 20% in length, or new ¹⁸F-FDG–positive metastases. With PERCIST, we propose a more than 30% increase in SUL peak, new ¹⁸F-FDG–avid lesions, or growth in lesion total lesion glycolysis by more than 75%—somewhat more stringent criteria for progression.

New ¹⁸F-FDG–avid lesions associated with the CT abnormality most consistent with tumor and clearly not due to inflammation or infection can be considered progression. New ¹⁸F-FDG–avid foci unassociated with a CT finding may well represent progression but should typically be verified by a follow-up PET/CT scan, or by another verification method 1 mo after their initial presentation (Fig. 5). Sometimes, however, verification will not occur anatomically, such as in lesions in bone marrow or in the spleen. RECIST 1.1 has addressed these issues to some extent. Progression in the lungs, particularly in the presence of potential inflammation or infection while a patient is on treatment, should be viewed with great caution, as discussed in the revised response criteria in lymphoma (32,33). New pulmonary infiltrates after treatment are often due to inflammation or infection and should be excluded before progressive disease is classified.

The extent of increase in ¹⁸F-FDG uptake required to represent progression is unclear. It is also unclear if an increase in SUL of over 30% in a single lesion is truly progression if the lesion is not the hottest. It may be difficult for the most intense lesion to increase in uptake over 30%, as the lesion may be performing glycolysis at a rate that is the maximum possible for its blood supply. Thus, growth in lesion size or total glycolytic volume potentially may be more indicative of progression than a rise in SUL peak in some settings. We have proposed a 30% increase in maximal SUL of the most intense lesion, with an SUL of more than 1.5 mean liver + 2 SDs as progression and an absolute increase in SUL peak of 0.8 units. However, it is probable that a 30% increase may not be achieved in all cases of progression. Rising 30% is probably easier in less glycolytically active lesions. If 5 lesions are assessed, the increase in glycolysis would need to be a 30% increase in the summed SUL peaks for the 5 most active lesions after treatment, versus the summed SUL peak of the 5 most active lesions before treatment.

For this reason, an increase of 75% in total lesion glycolysis for the most active tumor is proposed. This metric is reportedly more variable (at least the volume component) than is SUL peak (104). Total lesion glycolysis of the up to 5 target metabolic lesions is recommended at a minimum. It is possible that total lesion glycolysis of all lesions of sufficient intensity will be a better metric of progression than that of a single lesion. Methods for delineating lesions for total lesion glycolysis based on threshold values have been developed and are entering practice (Fig. 6). Thus, PERCIST 1.0 recommends that these data be collected as part of trials including PET for treatment response assessment. It may also be reasonable to collect SUVmax data for a single pixel, though these data are not used in response determinations as presently configured.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 6. 

(A) Patient with extensive non-Hodgkin lymphoma before treatment. Tumor with most intense ¹⁸F-FDG activity is in abdomen. Transverse images of easily measurable right axillary lymph node on CT are shown for convenience. (B) Commercial software tool (PET Volume Computed Assisted Reading; GE Healthcare) was used to localize foci of ¹⁸F-FDG uptake greater than mean liver SUL + 2 SDs of normal liver background (red). Manual intervention is required to separate normal ¹⁸F-FDG–avid foci, including brain, heart, and excreted urine, from relevant tumor. This semiautomated segmentation can be used to estimate total lesion glycolysis.

It is rare for an ¹⁸F-FDG–avid tumor to progress in the fashion of a tumor that is not ¹⁸F-FDG–avid, at least for measurable lesions. Small metastases, such as in the lungs, could be falsely PET-negative early in their progression. However anatomic progression that is not ¹⁸F-FDG–avid by RECIST or IWC in a previously ¹⁸F-FDG–avid tumor and that does not otherwise meet PERCIST criteria for progression would need verification before being considered progression.