Correlations between baseline ¹⁸F-FDG PET tumour parameters and circulating DNA in diffuse large B cell lymphoma and Hodgkin lymphoma

Thirty patients with diffuse large B cell lymphoma (DLBCL) and 50 patients with classical Hodgkin lymphoma (cHL) were enrolled prospectively. Description of the population with clinical characteristics and outcomes is available in two previous studies [15, 16]. All patients had performed a PET/CT on one of 3 different machines (description summarized in Additional file 1: Table 1). As one machine (PET/CT n°3) represented only 6,25% of the examinations performed (n = 3 for DLBCL and n = 2 for cHL), the patients concerned were excluded to avoid a bias linked to the PET/CT used. Description of the population studies, values of PET parameters and blood cell-free DNA measurements is available in Table 1. For the analysis concerning ctDNA, 3 patients for DLBCL and 15 patients for cHL were excluded as no mutation corresponding to lymphoma was found in ctDNA (false negatives of the technique at baseline).

Comparison of PET parameters distribution according to the machine use is summarized in Additional file 1: Table 1. No statistically significant difference (Wilcoxon–Mann–Whitney test p values > 0.05) was observed in the distribution of all PET parameters for DLBCL between PET/CT n°1 (n = 14) and PET/CT n°2 (n = 13). For cHL, only four parameters (TumBB, Dmax, itErosion, medPCD) among the twelve studied had a statistically significant difference (Wilcoxon–Mann–Whitney test p values < 0.05) in the distribution between PET/CT n°1 (n = 14) and PET/CT n°2 (n = 34). For these four parameters, a post-reconstruction harmonization by ComBat [17, 18] was performed to avoid a bias linked to PET/CT used (see Additional file 2: Fig. 1), with no statistically significant difference found between the distributions after harmonization (Wilcoxon–Mann–Whitney test p values > 0.05) and the results obtained used for the subsequent analysis.

Histograms of the distributions of each PET parameter, fused for DLBCL and cHL, are presented in Fig. 1. TMTV, TMTS, TumBB, Dmax and nROI showed non-different distributions for both DLCBL and cHL (Wilcoxon–Mann–Whitney test p values > 0.05), while the distributions were significantly different for the other parameters (Wilcoxon–Mann–Whitney test p values < 0.05).

Plots of high Spearman’s correlations (p < 0.05 controlled by Benjamini–Hochberg–BH correction [19]) between PET parameters for DLBCL and cHL are presented in Fig. 2. Three similar clusters were found in both DLBCL and cHL combining highly correlated parameters: the first one concerning the tumour burden with TMTV, TMTS and TLG, the second describing the tumour dispersion with TumBB and Dmax (and also with nROI with a lower correlation) and the third concerning the massiveness with medEdgeD and itErosion (but also TVSR and medPCD, which is, in the case of DLCBL, also associated with the tumour burden). MedEdgeD, itErosion, TumBB were also well correlated with TMTV and TLG for DLBCL. All statistically significant correlation values are visible in Additional file 3: Fig. 2.

Table 2 represents the Spearman’s correlations between the twelve PET parameters and the two blood DNA parameters ([cfDNA], [ctDNA]) for DLBCL and cHL, with the 3 first higher values parameters with an asterisk for each category (p values controlled by BH correction).

Concerning cfDNA and for DLBCL, three PET burden parameters (TMTS, TMTV and TLG) had the highest correlations with [cfDNA] (ρ = 0.78, ρ = 0.76 and ρ = 0.74, respectively). Parameters of dispersion, like TumBB, and of massiveness, like medPCD, were also well correlated (ρ = 0.69 and ρ = 0.70, respectively). For cHL, no statistically significant correlation was found between PET parameters and [cfDNA].

Concerning ctDNA, dispersion parameter TumBB had correlation values with [ctDNA] among the highest for both DLBCL (ρ = 0.83) and cHL (ρ = 0.54). For DLBCL, medPCD, describing massiveness and TMTS were also highly correlated with [ctDNA] (ρ = 0.81 for both). For cHL, two burden parameters, TLG and TMTV, presented the highest correlation values with [ctDNA] (ρ = 0.58 and ρ = 0.55, respectively).

In an univariate linear regression analysis (see Additional file 4: Table 2), most of the parameters were significantly associated with [cfDNA] of DLBCL and none to [cfDNA] of cHL. Concerning [ctDNA], burden parameters (TMTV, TMTS and TLG) were significantly associated with it in both DLBCL and cHL. Dispersion parameters (TumBB and Dmax) were significantly associated with [ctDNA] in cHL and massiveness parameters (itErosion, medPCD, medEdgeD) with [ctDNA] in DLBCL.

In the multivariate stepwise regression analysis used to determine [cfDNA] and [ctDNA] from PET parameters (see Table 3), statistically significant formulas were found for both DLBCL and cHL even if adjusted R² was higher for DLBCL (0.53 and 0.44 for [cfDNA] and [ctDNA], respectively) than for cHL (0.28 and 0.40 for [cfDNA] and [ctDNA], respectively). Interestingly, TMTS and medEdgeD were associated with determined [ctDNA] in both DLBCL and cHL, while TMTV was absent of the formulas used to predict [ctDNA].

Patients with DLBCL and cHL were enrolled prospectively in two non-interventional studies: LymphoSeq (NCT 02339805) and XPO1 (NCT 02815137). Clinical features and biological material at the time of diagnosis, including DNA from blood, were collected before any treatment. Patients were followed after rituximab-cyclophosphamide-doxorubicin-vincristine-prednisone (R-CHOP) or R-CHOP-like chemotherapies for DLBCL, or after ABVD or BEACOPP-escalated chemotherapies for cHL. An ¹⁸FDG-PET/CT was performed at the time of diagnosis and during the follow-up (mid-treatment and end of treatment, according to treatment strategies).

Patients provided written informed consent in accordance with the Declaration of Helsinki, and the Institutional Review Board of Henri Becquerel Cancer Centre approved the protocol (registration clinical.gov numbers: NCT02339805 and NCT02815137).

Blood samples were obtained by blood tests at diagnosis on EDTA tubes and were centrifuged for 10 min at 3000–3500 rpm within three hours of collection. Plasma was aliquoted into 1 mL in microtubes and stored at -80 °C until extraction.

Circulating cell-free DNA ([cfDNA]) was extracted from 3 mL of plasma aliquots with Amp Circulating Nucleic Acid® QI Kit (Qiagen, Hilden, Germany) according to the manufacturer's instructions and quantified using QuBit High Sensitivity dsDNA (ThermoFisher Scientific, Illkirch, France). After tumour sequencing adapted to the type of cancer [15, 16], the DNA was eluted in 60 to 80 μL of AVE buffer and then stored at − 80 °C. Quantification of the double-stranded DNA was performed by fluorometry on Qubit 2.0 (ThermoFisher Scientific Carlsbad, CA, USA), with Qubit® dsDNA kit HS Assay (ThermoFisher Scientific, Carlsbad, CA, USA). The circulating cell-free tumour DNA ([ctDNA]) concentrations were expressed in haploid genome equivalents per mL of plasma (hGE/mL) and calculated by multiplying the mean variant allelic frequency (VAF) for all mutations used for detection calling by the concentration of [cfDNA] (pg/mL of plasma) and dividing by 3.3, using the assumption that each haploid genomic equivalent weighs 3.3 pg, as previously described in the publication by Scherer et al. [28].

All patients underwent FDG PET/CT before the onset of chemotherapy, performed after a 6-h fasting and when blood glucose level was less than 1.7 g/L. PET data were acquired on 3 different PET systems, approximately 60 min after injection of 3.5 to 4.5 Mbq/kg, from the mid-thigh toward the base of the skull, 3 to 4 min per bed position. CT scan was set up to 100 to 120 kV, with an intensity modulation system. Injected activity, acquisition time and CT parameters were depended of the PET system and patient’s habitus. No contrast agent was administered to the patients for CT. Three different PET/CT scanners were used without selection according to patients: a Biograph16 HiRes (Siemens®, Germany) (PET/CT n°1), a GE710 (General Electrics®, USA) (PET/CT n°2) and a Biograph40 mCT (Siemens®, Germany) (PET/CT n°3) with acquisition parameters described in Additional file 1: Table 1. PET system was normalized daily and the calibration coefficient validated if the day-to-day variation remained below 0.3%. The global quantification, from the dose calibrator to the imaging system, was measured internally on a quarterly basis and double checked by the EARL’s quality assurance program. SUV was normalized according to the weight of the patients. Regions of interest were segmented semi-automatically with PET VCAR (General Electrics®, USA) based on 41% SUVmax segmentation, by consensus of two nuclear medicine physicians (SB and PD) with a visual control and manual adaptation if necessary. The spleen was considered as involved if there was focal uptake or diffuse uptake higher than 150% of the liver background. The bone marrow involvement was only included in the volume measurement if there was focal uptake. Contours (in RTSS format) were converted to binary mask (in mha format) by using the software plastimatch [29]. Two images mask was used for each patient: one binary mask with only contours, one 32-bits mask with SUV values hard-coded in the pixel data. PET parameters were extracted by using the in-house software Oncometer3D version 1.0.

The twelve following parameters were determined on the baseline PET/CT by an in-house software called “Oncometer3D”, with a graphical representation of the parameters visible in Fig. 3: (1) SUVmax, the highest maximal standardized uptake value (SUV) measured in all tumours, (2) SUVmean, the mean value of SUV measured in all tumours, (3) total metabolic tumour volume (TMTV) obtained by summing the metabolic volumes of all the nodal and extra-nodal lesions, (4) total lesion glycolysis (TLG) calculated as the product of the MTV and the SUVmean (TLG = TMTV × SUVmean), (5) total metabolic tumour surface (TMTS) obtained by summing the metabolic surfaces of all tumours [12], (6) tumour volume surface ratio (TVSR) corresponding to the ratio of the TMTV and the TMTS [12], (7) volume of the bounding box including the tumours (TumBB) corresponding to the volume of tumour dispersion, (8) the maximal tumour distance (Dmax) corresponding to the distance between the two lesions that were the furthest apart [20], (9) the number of regions of interest (nROI) corresponding to the number of unique tumour on the whole examination, (10) iterative erosion (itErosion) corresponding to the number of erosions [30] required to remove tumours from the images, (11) the median distance between the centroid of the tumours and the periphery (medPCD), (12) the median edge distance (medEdgeD) corresponding to the median distance between the opposite edges of the tumours.

In order to analyse a potential “machine effect” due to the use of several different PET/CT machines, the distributions of the different PET parameters according to the machines used and the type of disease were compared by Wilcoxon–Mann–Whitney tests. In case of differences in the distributions, a harmonization by ComBat method was carried out [17, 18].

To illustrate and compare distribution of PET parameters between DLBCL and cHL, fused histograms were produced with Wilcoxon–Mann–Whitney tests performed between the distributions of DLBCL and those of cHL.

To characterize the relationships between the different parameters, Spearman’s rank correlation coefficient of PET/CT parameters with each other and with DNA parameters ([cfDNA], [ctDNA]) was calculated separately for DLBC and cHL. For [ctDNA], null values were excluded from the analysis as they correspond to a false negative (mutation of the lymphoma not present in the panel and, therefore, no possibility to measure the ctDNA).

An univariate linear regression analysis was performed to explore linear correlations between PET and DNA parameters. To explore the interest of the combination of several factors, a multivariate stepwise regression analysis in both direction was used to determine [cfDNA] and [ctDNA] from the twelve parameters for both diseases.

Statistical significance was considered at p < 0.05 controlled by Benjamini–Hochberg correction [19]. All statistical analyses were performed using R software version 3.4.4 [31].