Defining the optimal method for measuring baseline metabolic tumour volume in diffuse large B cell lymphoma

Diffuse large B cell lymphoma (DLBCL) is the commonest subtype of lymphoma, representing 30% of lymphoid malignancies [1]. There has been a significant improvement in cure rates in recent years, with the addition of rituximab to cyclophosphamide, adriamycin, vincristine, and prednisone (CHOP) chemotherapy. However, a significant proportion of patients will progress or relapse after R-CHOP [2, 3] and long-term cure rates are only about 60% [4]. Whilst first line treatment has become more successful, salvage therapy after up-front rituximab has become less effective [5, 6]. It is important therefore to be able to reliably assess both pretreatment risk and identify patients at high risk of progression or relapse early to tailor treatment and test alternative approaches [7].

The International Prognostic Index (IPI) is currently used for estimating pretreatment risk, despite the fact that IPI often does not reliably predict individual patient outcome because DLBCL tends to behave heterogeneously [8]. Other factors that can predict prognosis, such as cell of origin or specific translocations, e.g. double-hit lymphoma (myc and bcl-2 translocations), have been identified but have not resulted in therapeutic advances as yet [9, 10].

The response to treatment in DLBCL has great prognostic value. Complete remission at the end of chemotherapy is associated with a high rate of progression-free survival (PFS) [11], but this information is obtained too late for choosing treatment. Positron emission tomography (PET) has been found to be useful in early monitoring of treatment for aggressive lymphomas [12]. In Hodgkin lymphoma, published multicentre trials support the use of early ‘interim’ PET for response-adapted treatment [13, 14]. However, in DLBCL, whilst initial reports suggested interim PET could reliably predict chemoresistance to CHOP [15, 16], later reports suggested the introduction of rituximab might affect the interpretation of “positive” interim PET scans [1, 17, 18]. Currently, the PFS of patients with a positive interim scan treated with R-CHOP is around 50% at 2–5 years [17, 18]. Attempts to standardise PET reporting [11, 19] and improve the positive predictive value of interim PET using semi-quantitative approaches [20] have not been sufficiently improved to enable interim PET to discriminate a group with poor prognosis in whom a change of treatment would be warranted [21, 22].

Baseline imaging characteristics can also predict outcome [23], including tumour burden [11]. The MInT study demonstrated a linear relationship between maximum tumour dimension and prognosis in patients treated with R-CHOP [4]. More recently metabolic tumour volume (MTV) has been identified as a promising baseline prognostic factor [11, 24, 25] that is superior to size-defined bulk [26, 27]. The high contrast afforded by 18F fluorodeoxyglucose (FDG) PET imaging may overcome some of the interobserver variability reported when segmenting tumour regions using computed tomography (CT) and it appears that PET is closer to the ‘ground truth’ when a tumour is delineated using PET compared to CT in solid tumours [28, 29]. The use of PET automatic delineation methods may also reduce interobserver variability [30].

Several methods have been proposed to measure MTV and applied in selected patients with large cell lymphoma. This has resulted in different cut-offs for MTV that separate good from poor prognostic groups [24‐26]. We recently reported our experience measuring MTV using software developed in-house. We combined baseline MTV with early response assessment using Deauville criteria in consecutive unselected patients with DLBCL treated with R-CHOP at a single institution [26] using quality assurance methods developed for clinical trials [31]. Using this approach, a third of patients were found to have high baseline MTV with incomplete early metabolic response after 2 cycles of R-CHOP and 5y-PFS of only 30% [26].

Validation of these data will require large patient numbers and involvement of international groups. Standardisation of the methodology for MTV is crucial for this endeavour, as previously occurred with the assessment of PET response using the Deauville criteria [11, 19]. Methods also need to be available using commercial software and be robust and easy to use in daily practice.

The aim of this study therefore was to:

Consecutive patients with DLBCL treated with R-CHOP at Guy’s and St Thomas’ NHS Trust from 2005 to 2012 were included [26]. Baseline PET/CT scans were acquired after a 6-h fast and 90 min after administration of FDG produced in an on-site cyclotron.

Images were acquired from the base of the skull to upper thighs using DST or VCT scanners (General Electric, Waukesha, WI, USA) for 5 minutes per bed position with separate head and neck views, if required. CT parameters were 140 kV; 115 mA; 0.5-s rotation time; 1.375 pitch. Images were reconstructed using iterative reconstruction and displayed using Hybrid Viewer (Hermes Medical Solutions, Sweden) scaled to a fixed standardised uptake value (SUV) of 10 and using a standard colour table.

MTV was measured on the baseline PET scan by one observer (HI) using:

The first three methods involved automatic segmentation of areas of tumour selected by the operator using a single-click for each region.

In the PERCIST method, the operator placed a 3-cm³ VOI in the right lobe of the liver. A wizard named ‘Tumour finder’ then automatically segmented all volumes within the image with uptake ≥1.5 x mean SUV + 2 standard deviations (SD) in the liver VOI. We also tested the exploratory threshold of 1 x mean SUV + 2 SD suggested [32], but found it to be too sensitive, selecting multiple areas that did not contain tumour (data not shown). If the liver showed extensive lymphoma involvement, a 1 × 1 × 2-cm VOI was placed in the descending thoracic aorta and used as the reference region instead [32].

The operator then modified volumes as required—manually removing regions that contained only physiological FDG uptake, e.g. brain or bladder, or by using editing tools to remove physiological uptake adjacent to the tumour that had been automatically included in the volume, e.g. myocardial or urinary tract and bowel uptake.

Individual tumour volumes, where more than one volume was present, were summed to calculate the total MTV. Observers were blinded to patient outcome.

Baseline MTV, using FDG-PET, is a promising prognostic indicator in patients with DLBCL, which is better than using size-defined bulk [25, 26]. Tumour lesion glycolysis, which is the MTV multiplied by the mean SUV in the volume, is also prognostic [37], but appears no better than MTV in DLBCL [26, 27]. Cut-offs ranging from 220 to 600 cm³ have been reported to separate patients into groups with low and high baseline MTVs (Table 3) which are predictive of PFS and OS. Cut-offs have been derived using ROC curve analyses [24‐27] that depend on the distribution of values in the dataset, which are influenced by patient characteristics (Table 3), with populations with worse clinical characteristics tending to have a higher optimal cut-off for MTV, but also crucially, as demonstrated in our study, on the method used to outline the tumour volume. The influence of the method of measurement on the optimal cut-off has been previously reported in 59 patients with Hodgkin lymphoma [39] and 106 patients with T cell lymphoma [40]. For clinical use, a consensus will be required on a suitable method and an optimal cut-off to define the MTV for specific lymphoma subtypes and treatment regimens, which will require validation in multicentre prospective trials.

Algorithms have already been developed for segmentation of volumes for radiotherapy planning purposes in solid tumours [41, 42]. Boundaries can be chosen using an absolute SUV value or a percentage of the maximum SUV. Alternatively, more complex methods may be adopted, such as contrast-orientated, possibility theory and adaptive thresholding. No single method is likely to perform optimally in every patient, and consensus methods, such as the majority vote, have been reported to improve accuracy compared with the ‘ground truth’ of manual delineation by experts or surgical specimens [43]. In a recent publication, consensus methods performed better than the worse performing of three established automatic segmentation methods and were close to the best-performing method in all patients [43]. Five segmentations were implemented in a single software platform for evaluating patients scanned on four different cameras with lung and breast tumours and which also included eight patients with lymphoma.

So far in DLBCL, three methods have been proposed in the literature for measuring MTV [26, 32, 33]. Importantly none of these methods are vendor-specific and we have demonstrated that measurement using the 2.5 method is robust using in-house software and commercial software. Efforts are being made to develop automated freeware incorporating all the published methods [39], but whether this will be acceptable for making patient management decisions using MTV as a prognostic tool remains to be seen.

We tested these methods in a population of consecutive patients with de novo DLBCL treated with standard R-CHOP at a single institution, likely to be representative of the general patient population. We did not measure CT-based tumour burden as the CT component of the PET-CT scans were performed as low-dose non-contrast scans, in keeping with our usual clinical practice. The first method measured any activity that may be significant with a SUV greater than 2.5 [26]. The second method was derived from phantom experiments to give the best estimate of anatomical volume [33]. The third method also measured any significant activity, but using liver uptake as the threshold [32], which may be less influenced by factors that cause inaccuracy in SUV measurement, but which may be more dependent on patient preparation and metabolic status, with reduction in normal liver uptake observed when there is very high tumour burden at baseline [38].

The in-house and commercial methods for measuring MTV using the 2.5 method gave almost identical results. The PERCIST method was very close to the 2.5 method, but probably overestimated MTV in approximately 12% of patients who had low FDG uptake in the liver or liver involvement by lymphoma. The 41% method was very different in absolute MTV values compared to the other methods and was more susceptible to measurement variability when there was tumour heterogeneity.

Accordingly, we found the optimal cut-off for MTV to predict PFS ranged from 166 to 400 cm³. Although all three methods could predict PFS with similar accuracy in the overall study population, we found for some individual patients with very intense masses, the 41% method appeared to underestimate tumour volume compared with the other methods. The 2.5 method gave an optimal cut-off in our study population (400 cm³) which was in the middle of the cut-offs previously reported by Song and colleagues using this method in two publications. The first measured MTV in good-prognosis patients with no extranodal involvement (derived cut-off 220 cm³) [25] and the second in poor-prognosis patients, all of whom had bone marrow involvement (derived cut-off 600 cm³) [44]. The 41% method gave an optimal cut-off in our population which was much lower than the 550 cm³ [24] and 300 cm³ [40], respectively, reported by Meignan and colleagues in two publications. There were twice as many patients over the age of 60 in the study with the lower cut-off, which is surprising as increased age is generally associated with worse prognosis. Therefore, the cut-off might have been expected to be higher in an older population (Table 3). Other clinical characteristics were similar in these two studies. The variability in the cut-offs reported for the 41% method raise concerns that the optimal cut-off may be more dependent on how regions are selected by different groups, when there is considerable tumour heterogeneity.

There was high interobserver agreement for measuring MTV with all methods. The 41% method was the most complex to use in our experience, reflected in the time taken to measure MTV in a subset of 50 patients. The PERCIST method was usually the quickest, because it allowed automatic segmentation of all regions on the scan, using the ‘tumour finder’ wizard (©Hermes Medical Solutions). This was despite needing to edit out areas that had uptake above the liver, accounted for by areas with high physiological uptake such as the brain and bladder. Inflammatory uptake might also require editing, but we did not observe this in the patients in our study (Fig. 6).

The study confirmed that the prognostic role of baseline MTV [26] using software developed in-house could be reproduced accurately using commercially developed software. We previously found that baseline MTV was a good prognostic indicator, better than size-defined bulk [26]. Using all three methods, 5y-PFS in the current study was similar to the values reported in our earlier manuscript of 43% for patients with high MTV compared to 85% for patients with low MTV [26] and compares favourably with previous publications [24, 40]. The sensitivity and specificity of the three methods is shown in Fig. 4.

We combined baseline MTV with early response assessment at two cycles in an attempt to improve prognostic value in our previous work [26]. High MTV and failure to achieve a complete metabolic response (Deauville score 4,5) at 2 cycles was found in 31% of patients who experienced 58% of study events with 5y-PFS of only 30%. Combining MTV with baseline factors rather than early response might be a more attractive option. Recently, El-Galaly and colleagues [23] reported that combining baseline PET findings with the new National Comprehensive Cancer Network (NCCN) IPI, which splits patients according to age groups >40, >60 and >75 years and by LDH levels 1–3 or >3 times the upper limit of normal, was better at predicting prognosis than PET combined with the IPI or revised IPI. Furthermore, the number of involved extranodal sites and the presence of bone/bone marrow, pleura and female genital organ involvement was associated with inferior prognosis. Combining NCCN-IPI and MTV and possibly other baseline imaging features as suggested by el-Galaly and colleagues [23] might be even more informative.

The cell of origin is also known to influence prognosis, with the activated B cell (ABC) subtype conferring a worse prognosis than the germinal centre B cell (GCB) subtype [9]. Genetic rearrangements including overexpression of BCL2 and MYC which regulate apoptosis and proliferation area are also associated with inferior prognosis [10, 45]. Cottereau et al. reported on 81 patients [40] mostly with advanced stage DLBCL, combining molecular profiling data with MTV. High MTV using a cut-off of 300 cm³ (by 41% method) was associated with identical 5y-PFS of 43% in our study. The subset of 16 patients with high MTV and the ABC subtype had 5y-PFS of 30% and OS of 23%. Patients with overexpression of BCL2 and/or MYC had inferior prognosis irrespective of MTV. MTV, however, separated the 55 remaining patients into good- and intermediate-prognosis groups. This suggests a potential for the strategy of combining imaging and other biomarkers for pretreatment risk referred to as ‘Radio(gen)omics’. Evaluation will involve pooling of data to derive and validate risk estimates with international collaboration.

In summary, all the published methods for measuring MTV in DLBCL were prognostic in our study for PFS and OS. The optimal cut-off using the 2.5 method in this unselected patient population was in line with cut-offs published by another group using this method in two populations with good and poor prognosis, respectively. A limitation was that scans were acquired at 90 min, longer than currently recommended by EANM procedural guidelines [46]; nonetheless, in our hands, the 2.5 method had the advantage of being easy to use and reproducible across different software platforms and between observers. In our opinion, contouring methods based on percentages of the maximum uptake in the volume may be easier to apply in solid tumours [42, 46] than in DLBCL, where patients often present with multiple regions with heterogeneous uptake. Developments in software may overcome some of the difficulties with measurement that we encountered.

The methodology is evolving and will require prospective validation in sufficiently large patient cohorts combined with other prognostic factors, to determine whether robust pre-treatment risk estimates can be identified to select patients in whom to test alternative treatments including novel agents.