KRAS mutation effects on the 2-[18F]FDG PET uptake of colorectal adenocarcinoma metastases in the liver

Background Deriving individual tumor genomic characteristics from patient imaging analysis is desirable. We explore the predictive value of 2-[18F]FDG uptake with regard to the KRAS mutational status of colorectal adenocarcinoma liver metastases (CLM). Methods 2-[18F]FDG PET/CT images, surgical pathology and molecular diagnostic reports of 37 patients who underwent PET/CT-guided biopsy of CLM were reviewed under an IRB-approved retrospective research protocol. Sixty CLM in 39 interventional PET scans of the 37 patients were segmented using two different auto-segmentation tools implemented in different commercially available software packages. PET standard uptake values (SUV) were corrected for: (1) partial volume effect (PVE) using cold wall-corrected contrast recovery coefficients derived from phantom spheres with variable diameter and (2) variability of arterial tracer supply and variability of uptake time after injection until start of PET scan derived from the tumor-to-blood standard uptake ratio (SUR) approach. The correlations between the KRAS mutational status and the mean, peak and maximum SUV were investigated using Student’s t test, Wilcoxon rank sum test with continuity correction, logistic regression and receiver operation characteristic (ROC) analysis. These correlation analyses were also performed for the ratios of the mean, peak and maximum tumor uptake to the mean blood activity concentration at the time of scan: SURMEAN, SURPEAK and SURMAX, respectively. Results Fifteen patients harbored KRAS missense mutations (KRAS+), while another 3 harbored KRAS gene amplification. For 31 lesions, the mutational status was derived from the PET/CT-guided biopsy. The Student’s t test p values for separating KRAS mutant cases decreased after applying PVE correction to all uptake metrics of each lesion and when applying correction for uptake time variability to the SUR metrics. The observed correlations were strongest when both corrections were applied to SURMAX and when the patients harboring gene amplification were grouped with the wild type: p ≤ 0.001; ROC area under the curve = 0.77 and 0.75 for the two different segmentations, respectively, with a mean specificity of 0.69 and sensitivity of 0.85. Conclusion The correlations observed after applying the described corrections show potential for assigning probabilities for the KRAS missense mutation status in CLM using 2-[18F]FDG PET images.

regarding proliferation rate, differentiation and heterogeneity, or genomic profile of malignant tumors [3,4] potentially improves selection of the appropriate treatment pathways for targeted therapies while minimizing the need and risks of interventional procedures and maximizing patient comfort [5].
Colorectal cancer (CRC) patients with Kristen rat sarcoma viral gene (KRAS) mutant tumors are associated with lack of response to anti-epidermal growth factor receptor (anti-EGFR) antibody therapy [6][7][8][9][10]. It has been also shown that KRAS mutations are a significant predictor of overall survival in metastatic CRC and of recurrence after surgery or radiofrequency ablation of colorectal cancer liver metastases (CLM) [11,12]. KRAS mutation has been associated with relatively higher rates of local failure or positive resection margins after thermal ablation or resection of CLM [11,12]. Recently, others have found that the coexistence of KRAS mutation with increased 2-[18F]FDG uptake is a negative prognostic factor in primary CRC [13]. KRAS mutation status has a predictive value also for image-guided ablation for lung adenocarcinoma [14].
While primary and metastatic CRC sites have a high concordance for KRAS mutations, meta-analyses suggest the concordance may not be 100%. An initial high (100%) KRAS genomic concordance has been observed between the primary tumor and secondary lesions in metastatic CRC [15,16]. A subsequent meta-analysis showed that the level of concordance may vary, reporting an overall concordance of 94.1% between 986 pairs of primary and distant metastasis from 17 publications [16]. In a more recent review, Mao et al. [17] have observed a pooled concordance rate of 92%. These data suggest that occasionally KRAS mutant primary tumors harbored KRAS wild-type metastasis and, as such, could potentially benefit from anti-EGFR treatment [17]. Specifically for CLM, these authors found 8.0% false positive (wild-type or normal expression in primary tumor but mutant or loss of expression in metastases) and 9.7% false negative (mutant or loss of expression in primary tumor but wild-type or normal expression in metastases) rates [17]. Additionally, patients treated with targeted therapies, including EGFR, HER2 and BRAF targeting combinations, have been reported to develop KRAS mutations at resistance, often in a heterogeneous pattern involving some lesions [18,19]. Deriving the genomic properties through metabolic imaging of individual lesions may optimize subsequent interventions (Fig. 1).
KRAS mutations appear to be related to the glucose metabolism [20]. Therefore, KRAS mutant lesions may be distinguishable using images of metabolic activity. The manifestation of KRAS mutations in PET images has been studied for non-small cell lung cancer (NSCLC) patients [21,22] as well as for CRC [13,[23][24][25][26][27][28][29][30]. For NSCLC, one study showed that a multivariate model, including 2-[18F]FDG SUV MEAN , may be used as a predictive marker for KRAS mutations [21]. However, others found that EGFR mutations correlate with NSCLC PET image features, whereas KRAS mutations do not [22]. Similarly for CRC, one investigation of the effect of KRAS mutations on 2-[18F]FDG uptake of CRC lesions found no correlation [23], while others found statistically significant correlations for primary tumors [13,[24][25][26]. Several SUV features including SUV MAX and SUV MEAN were found to significantly correlate with the presence of KRAS mutations for newly diagnosed rectal cancer [27], although predictive value of the 2-[18F]FDG uptake was low. Other investigations showed statistically significant separation of the KRAS-mutated tumors by SUV MAX and SUV PEAK when a C-reactive protein (CRP) upper threshold of 6 mg/L was introduced [28], or when SUV MAX was used in a decision tree also utilizing CT texture and blood flow, derived from dynamic contrast-enhanced CT [29]. For CRC metastases, SUV MAX was significantly associated with KRAS mutations when considering only tumors larger than 10 mm in diameter in order to minimize bias due to the partial volume effect (PVE) [30]. Primary tumor SUV MAX was found to be significantly higher in CRC subjects with KRAS mutation compared to wild type [13]. None of the above studies were specific to CRC liver metastases (CLM).
We focus on evaluating KRAS effects on 2-[18F]FDG uptake specifically for CLM. In addition, we apply corrections to the 2-[18F]FDG uptake for: (1) partial volume effect and (2) time of tracer uptake from injection to PET scan acquisition. The latter is important for the investigated dataset which contains only PET/CT images obtained in the interventional setting during PET/CTguided ablations or biopsies for which the uptake time can vary. Biopsies and ablations performed under realtime PET/CT guidance are useful to target tumors with poor conspicuity in non-functional/anatomic only imaging [31][32][33][34][35]. If the specimen from a biopsy during such a procedure is subjected to genomic profiling, the molecular and metabolic data are obtained simultaneously for the same lesion, thereby removing uncertainty for the molecular status of the imaged lesion.

Patients
The records of thirty-seven (37) patients who underwent PET/CT-guided biopsies of colorectal adenocarcinoma liver metastases (CLM) in the period between April 2011 and June 2019 with molecular pathology reports were reviewed under an Institutional Review Board-approved retrospective research protocol. Thirty-nine PET/CT scans were analyzed since two patients underwent a second PET/CT-guided biopsy for new liver lesions, which were also included in the analysis. Altogether 60 CLM were segmented in these 39 interventional PET scans using two different PET segmentation tools. Twelve of the included lesions were in the vicinity of previously treated (ablated or resected) region of the liver.
The KRAS mutational status for each case was extracted from the molecular pathology report of each case. Thirty-one (31) of the 60 lesions had mutational status assigned from the PET/CT-guided procedure. Other lesions in the same cases as well as cases for which the mutational status was known from another specimen were kept in since the probability for mislabeling is small. The investigated dataset is a compromise between reducing the number of potentially mislabeled cases (from about 10% to about 5%) and having a sufficient number of cases for the analysis.

PET/CT-guided biopsies
In PET/CT-guided biopsy procedures, after the initial PET/CT scan, the interventional radiologist places the needle by targeting the PET-avid region of the lesion after which a second CT scan with the needle in place is performed and fused with the pre-insertion PET, PET pre , for verification of the needle placement [33]. In cases where mis-registration with PET pre is suspected due to motion, a second real-time PET with the needle in place may be obtained [36].
All PET/CT-guided biopsies were performed at the same interventional PET/CT scanner (Discovery 690, GE Healthcare, Waukesha, Wisconsin). The average injected 2-[18F]FDG activity (mean ± SD) was 152.0 ± 52.4 MBq (4.1 ± 1.4 mCi). The post-injection residual activity in the syringe was measured only for one protocol case (6.2 MBq for case # 2), while a residual activity of 11.1 MBq (0.3 mCi) was assumed for all other cases based on our current clinical procedures. Uptake times varied from 40.7 to 205.7 min (mean 83.8 min, std 34 min) due to interventional work specifics: Optimizing the flow of the interventions and patient safety and emergency procedures have a higher priority.
PET scans were centered in the liver and obtained for one or two bed positions consisting of 47 transverse slices, 3.27 mm thick with 11-slice overlap. Transaxial field of view of 70 cm and 128 × 128 image matrix (5.47 × 5.47 mm pixels) were used for all cases. The acquisition time typically varied from 3 to 5 min per bed for the pre-needle insertion PET scans included in this study. PET scans with breath hold [33] were excluded since their number was very small and breath hold affects the SUV. The reconstruction settings were the same for all scans (time-of-flight on, two iterations, 16 subsets, sharpIR system modeling, post-filter: 6.4 mm, axial filter: heavy).

Molecular pathology
Molecular pathology reports were obtained from the patients' medical record and reviewed for all 37 patients. If the patients had more than one pathology report, the report from the PET/CT-guided biopsy of the targeted CLM or, if not available, the report closest in time to that biopsy was used. KRAS mutation status was determined by clinically validated molecular assays including the mass spectrometry-based Sequenom technologies [37] test only (n = 5) and the next-generation targeted exome sequencing by Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT) [38] with or without Sequenom (n = 32). In addition to KRAS missense mutations, MSK-IMPACT also detects KRAS copy number alterations such as whole gene amplifications. The KRAS mutation status for each case was noted but was not provided to the operators performing tumor segmentation.

PET image analysis
Sixty CLM were segmented using the PET-edge tool of MIM (MIM Software, Inc., Cleveland, Ohio) and a fixed SUV threshold in Hermes Gold LX (Hermes Medical Solutions, Stockholm, Sweden). Two different researchers performed the segmentations using the two packages and were blinded to the patient mutational status and to each others' results. All automatically generated contours were visually inspected, and small manual corrections were applied if needed. In Hermes, a threshold of SUV = 4.0 or higher was used: The segmentation thresholds for three higher background cases were increased to avoid unrealistically large volumes.
Mean and peak standard uptake values (SUV, normalized to patient weight) as obtained from the two segmentation methods and the maximum SUVs obtained in Hermes were recorded for all target CLM. When SUV PEAK was not available due to a small CLM volume (37 cases out of the 60 lesions in MIM), SUV MEAN was used as an approximation.
The tumor-to-blood SUV ratio approach developed by van den Hoff et al.
[39] that corrects for variable uptake time from the FDG injection to the PET scan acquisition, variable tracer supply between subjects and technical uncertainty factors [40] was implemented. To determine the 2-[18F]FDG uptake in the blood, we manually segmented the part of the descending aorta visible in the PET images using Hermes. The contours were drawn at least 5 mm away from the edge of the aorta to minimize PVE effect on SUV mean (Fig. 2).

Partial volume effect correction
The limited PET resolution causes a partial volume effect which results in loss of accuracy in recovering the true activity especially in small objects. Since at PET resolution most small liver tumors can be approximated by equal volume spheres, an approximate partial volume effect correction can be applied by using the percent contrast (Q) for different-diameter higher activity spheres placed in uniform background as measured according to the NEMA 2.0 protocol for PET scanner acceptance [41]. During PET acceptance, the measurement is performed at a sphere-to-background activity concentration ratio SBR = 4:1. For our purpose, we filled all spheres of the same phantom with SBR = 2.19:1 (9.284 and 4.240 kBq/ mL, respectively) to approximate the mean tumor-tonormal liver ratio in the 39 patient scans. The lung insert in the center of the phantom was also in place with no activity. The phantom was scanned on the same PET scanner where all patients were scanned with two bed positions for 15 min each to reduce image noise. Spherical VOIs with diameters matching those of the inner diameter of the phantom spheres (10, 13, 17, 22, 28 and 37 mm) were centered in the CT images of the phantom spheres and then copied to the same location in the registered PET images using the Hermes software. Then, the mean, peak and maximum SUV for each sphere were recorded to calculate the respective recovery coefficients as described below.
The In analogy to (1), recovery coefficients can be defined also for the maximum and for the peak measured activities: which will allow to apply partial volume correction to SUV PEAK and SUV MAX using: The measured recovery coefficients for SUV MEAN , SUV PEAK and SUV MAX (Eqs. 1 and 3) are plotted for each sphere in Fig. 3. Corrections to these recovery coefficients were applied for the 1 mm thickness of the cold sphere walls using a RC model obtained by convolution of the resolution PSF with the spheres given the known sphere-to-background contrast (lines in Fig. 3) [42]. The cold wall correction was obtained from the ratio of the simulated recovery coefficients for the max, peak and mean SUVs for spheres with the same inner diameters with and without 1 mm walls, respectively. The sphere external diameters were confirmed by caliper measurements, and the wall thickness was verified using a micro-CT scan for some of the spheres. In the simulations, we performed convolution of a symmetrical Gaussian point-spread function with each sphere and used a sphere-to-background ratio of 2.26, which is close to the midpoint between the ratios in the NEMA phantom and the patients. The cold wall corrections to the RC ranged up to ~ 11.1% and 12.8% for 1 mL lesions for SUV MEAN and SUV MAX and were less than 5% and 0.2% for volumes larger than 10 mL, respectively. For extending the PVE corrections to the few lesions with volumes larger than the 37-mm-diameter sphere (26.52 mL), we followed the RC trends provided by the convolution-based model described above [42]. Note that according to our definition (Eq. 3), the RC for SUV MAX and SUV PEAK are larger than 1.0 for large lesions due to statistical effects (positive bias of maximum relative to mean).
The approximate SUV MEAN , SUV PEAK and SUV MAX PVE correction for each lesion with volume V les , RC eff (V les ) was obtained by interpolation of the final RC curves obtained after cold wall correction (lines in Fig. 3). The The final recovery coefficients used for PVE correction (lines) are obtained by applying a correction for the cold walls of the phantom spheres obtained by using a RC model [42] to compute the recovery coefficient ratio for spheres with and without walls mean background SUV for each lesion, SUV B,lesion , was measured in two PET slices using doughnut-shaped ROIs manually drawn to avoid the visible spill out from the respective lesion. The PVE-corrected SUV MEAN , SUV PEAK and SUV MAX for each lesion are then obtained using SUV B,lesion and Eqs. (2) and (4). For lesions close to the periphery of the liver, SUV B,lesion was less than the mean normal liver SUV, SUV B,mean liver . The latter, SUV B,mean liver , was obtained from large manually drawn ROIs in normal liver far from the lesions and was used for calculating the tumor-to-liver ratio, SUVTLR = SUV PVEC, lesion /SUV B, mean liver , for each lesion.

Uptake time correction
We employed the uptake time correction method developed by van den Hoff et al. [39] for CRC liver metastases and validated also for other lesions. For calculating the tumor-to-blood standard uptake ratio (SUR) employed by this method, we derived the blood SUV from the 2-[18F]FDG activity in the descending aorta. Since for liver interventions the PET scan is limited to two bed positions (83 slices) providing an axial field of view (FOV) from the inferior to the superior border of the liver, and we extracted SUV blood mean from an aorta volume at least 0.5 cm away from the aorta surface to minimize PVE, the aorta ROI volumes were smaller (average analyzed volume 1.8 mL) than those previously used [39]. The average SUV values for the descending aorta were 2.2 + − 0.6 mL/g. The injection and scan times were automatically extracted from the image headers using an in-house developed tool (DBbrowser). SURs were calculated as SUR MEAN = SUVlesion mean /SUV blood mean , SUR PEAK = SUV lesion peak / SUV blood mean and SUR MAX = SUV lesion max /SUV blood mean , using both MIM and Hermes segmentation contours of the lesions and Hermes contours for SUV blood mean . A correction for the difference between actual uptake time post-injection, T, and standard uptake time T 0 = 60 min was applied to each lesion's SUR MEAN , SUR PEAK and SUR MAX using the approximations to equations (8) and (10) proposed by van der Hoff et al.
with apparent volume of distribution, V r = 0.53, and parameter b = 0.313 as determined previously [39]. By reducing variance, this approach has led to improved correlations in several studies with similar and larger patient cohorts for different cancer types [43][44][45][46][47]. In one of these studies, which used 90 cases with dual time point 2-[18F]FDG PET, the uptake time varied from 56.4 to 197 min [44].
The lesion volume and the SUV mean, peak and max metrics were extracted immediately after segmentation. Then PVE correction was applied to these values according to the lesion volume. SUVTLR and SUR were calculated using the PVE-corrected mean, peak and max SUV, which was then followed by applying the uptake time correction prior to investigating the correlations.

Patient demographics and KRAS mutations
The patient demographics are given in Table 1. KRAS mutations were observed in 15 (37.8%) and copy number variations in 3 (8.1%) of the 37 patients, respectively (Table 2). In 16 of the 18 patients with KRAS mutant tumors (except #9 and #23 in Table 2) and 15 of the 19 patients with KRAS wild-type tumors, sequencing was performed on the target CLM from the same PET/CTguided biopsy. Of the 15 missense mutations, eight are in codon G12, four are in codon G13D, two in codon Q61H and one in codon A146T.
In addition to patient #11 (Table 2), two more patients have developed a KRAS mutation in codon Q61H after initial testing. Case #1 had no KRAS mutation in the primary and later, at the time of CLM targeting PET/ CT-guided biopsy, but showed a Q61H exon 3 KRAS mutation in a different lesion in a remote section of the liver 2.5 years later during preparation for radioembolization. Similarly, case #31 had no KRAS mutation in an older sigmoid rectum specimen (primary) but exhibited also a codon Q61H exon 3 KRAS mutation in the targeted CLM at time of the PET/CT-guided biopsy 5 years later. Altogether, both primary tumor and a metastasis were genotyped in 15 cases and the KRAS status was concordant in 13/15 (87%) cases.

PET image analysis
The average tumor volumes obtained with the PET-edge algorithm in MIM and with the SUV threshold algorithm in Hermes are 5.3 mL (min. 0.3, max. 31.3, std. 6.9 mL) Biopsy only and 6.1 mL (min. 0.2, max. 43.0, std. 8.5 mL), respectively. The average lesion diameter from the two segmentations is 22 mm. Before PVE correction, the mean lesion SUV of all 60 lesions was 6.2 ± 2.2 (MIM) and 6.0 ± 1.6 (Hermes) and the mean normal liver SUV is 2.7 ± 0.6 resulting in mean tumor-to-liver ratio of 2.2 and 2.3 for Hermes and MIM lesion contours, respectively. The mean background liver SUV measured in the vicinity of the lesions, SUV B,lesion , and used for the RC calculation [Eqs. (2) and (4)] is 2.4 ± 0.7. The average SUV values obtained for the descending aorta are 2.2 ± 0.6. The values for the different uptake metrics corresponding to the two segmentation methods for each mutation status are given in Table 3.

Correlations
Since only KRAS missense mutations were reliably shown to affect the outcome of anti-EGFR therapy, below we present the results when KRAS gene amplification cases are grouped with wild type. The results when the three cases carrying KRAS gene amplification and no KRAS missense mutations are grouped with KRAS missense mutations are presented in the supplement. The correlations observed between the investigated SUV metrics and the KRAS mutational status are presented in Table 4 using Student's t test and Wilcoxon rank sum test with continuity correction p values and area under the curve (AUC) of the receiver operating characteristic (ROC). The metrics presented in Table 4 are ordered with decreasing p values from top to bottom. This correlates with the increase in the AUC values which are presented only for the bottom row and the rightmost column.
The mean coefficient of variation of the AUC between the two segmentation methods for SUR MEAN , SUR PEAK and SUR MAX with both PVE and uptake time corrections is 4.3, 1.5 and 1.2% (from the AUC values for SUR PVE+time corr in Table 4), respectively. ROC curves and AUC for SUR MAX for KRAS mutation prediction

Table 3 Average 2-[18F]FDG SUV metrics (± 1 STD) for KRAS wild-type and missense mutated CLM obtained with two independent volume segmentations (MIM and HERMES)
KRAS gene amplification cases (n = 3) are grouped with wild type. Average metrics values after PVE and uptake time correction are provided for the derivative metrics tumor-to-liver ratio (SUVTLR) and tumor-to-blood ratio (SUR) as well are shown in Fig. 4. The sensitivities and specificities for predicting KRAS mutations by the SUR-derived metrics are shown in Table 5. Logistic regression curves for SUV MAX with and without PVE and uptake time corrections and for SUR MAX with both corrections are shown in Fig. 5. We see some dependence of the logistics regression optimal cutoff values, on the segmentation methods used. The optimal cutoff values for SUR after PVE and time correction for MIM versus Hermes segmentation are: 3.04 versus 3.11 for SUR MEAN , 2.89 versus 3.11 for SUR PEAK and 4.09 versus 3.78 for SUR MAX. The optimal cutoff however did not change between the two types of grouping for the KRAS gene amplification cases: with wild type as presented in

Table 4 Statistics for the separation of KRAS missense mutations based on 2-[18F]FDG uptake when KRAS gene amplification is grouped with wild type for all 60 lesions in 39 interventional PET/CT scans
Student's t test p values are given for all metrics, while Wilcoxon rank sum test with continuity correction p values (abbreviated to Wilc. R. Sum) and AUC are given as a second and third line number only for PVE-and time-corrected SUR (bottom row) and for the maximum values derived from the Hermes segmentation contours. Values obtained after PVE and uptake time correction as well as for derivative metrics as tumor-to-liver ratio (SUVTLR) and tumor-to-blood ratio (SUR) are provided  In 31/39 interventional PET/CT scans analyzed, the mutational status of the targeted lesion was determined from the specimen extracted from the same procedure. Applying the analysis for SUV MAX only to the 31 lesions with accurately known KRAS mutational status also showed separation of the two groups when two remaining amplification cases are grouped with the wild type (average Student's t test p values between the two segmentations: 0.003 for SUV MAX with no corrections, 0.020 after both corrections and 0.011 for SUR MAX after both corrections).
If we remove the 12 lesions in the vicinity of previously treated parts of the liver and analyze the remaining 48 lesions, the Student's t test p values for separating KRAS+ mutations from wild type and whole gene amplification based on maximum uptake are: 0.015 for SUV, 0.006 for SUV PVE , 0.008 for SUVTLR, 0.005 for SUV PVE+time cor. , 0.005 for SUR time cor. and 0.001 for SUR PVE+time cor. In this case, the mean AUCs between the two segmentations are 0.74 for SUV MAX, PVE + Time corr and 0.77 for SUR MAX,PVE+time cor . If within this group of 48 lesions KRAS whole gene amplification cases are grouped with KRAS+, statistical significance is lost.
When the three KRAS gene amplification cases are grouped with KRAS missense mutations, the p values are higher and the specificity and sensitivity and the AUC values are lower (see Additional file 1: Tables S1, S2 and Fig. S1).

Discussion
For the investigated patient cohort, applying PVE corrections to the SUV leads to better separation of the KRAS mutations for all three SUV-and SUR-based metrics: mean, peak and max (Tables 3 and 4

, Figs. 4 and 5).
Applying uptake time correction leads to noticeably better separation of the two groups only for the SUR metrics but not for the SUV metrics (Table 4). PVE-corrected tumor-to-liver ratios, SUVTLR, performed similarly to PVE-corrected SUV. Using the uptake time-corrected tumor-to-blood ratio, SUR, both with and without PVEC result in better separation of the two groups than PVEand uptake time-corrected SUV. This indicates that uptake time variability is not the only relevant factor affecting the correlation between the chosen uptake metric and the tumor's metabolic rate. The removal of inaccuracies related to injected activity due to the residual activity assumption, scanner calibration and body mass/  weight as well as accounting for inter-subject variability of the arterial tracer supply as achieved by the use of the SUR metric is of key importance as well [40]. In fact, the logistic regression analysis (Fig. 5) shows that for the current patient cohort only the PVE-and uptake time-corrected SUR MAX provides sufficient separation of the two groups to make mutation predictions from the 2-[18F] FDG uptake in PET images.
Using SUR MAX and applying the corrections described above seem to have allowed for better separation of KRAS mutated from wild-type CRC liver lesions than that demonstrated previously [24][25][26][27]30] (Table 4). When KRAS gene amplification is grouped with wild type, the PVE-and uptake time-corrected SUR MAX gives the highest predictive value: Between the two segmentations, the mean specificity is 0.694 and the mean sensitivity is 0.854 (sum 1.548, Table 5). For SUR MEAN and SUR PEAK , the mean specificity + sensitivity is slightly lower: 1.51 and 1.52.
Another factor we consider helpful in achieving better separation of the KRAS-mutated cases is an increased accuracy in labeling the mutational status. In 31/39 interventional PET/CT scans analyzed, the mutational status of the targeted lesion was determined from the specimen extracted from the same procedure. According to previous publications, about 6-10% of the remaining 29 lesions may have mutational status different from that determined in other lesions of the same patient [17]. Assuming that the overall discordance rate is approximately 10% [17], and since the mutational status of 31/60 lesions is precisely known, only 5% (0.1 *(60-31) = 2.9) of the 60 lesions may be potentially mislabeled. This decreases in half the general discordance rate. Thus, including target tumors with a precisely known mutational status (52% in this cohort) increases the accuracy of the dataset and may have contributed to better separation of the two mutations.
In our dataset, the KRAS missense mutations seem better separated (lower p values, higher AUC in Table 4 and Fig. 4) when the cases with KRAS gene amplification (3 of the 32 cases tested by MSK-IMPACT) are grouped with the wild type. This is understandable since KRASmutated cells were shown to exhibit enhanced glucose uptake [20] and these missense mutations were also found to be mutually exclusive with KRAS whole gene amplification in this as well as in other patient cohorts [48]. The optimal cutoff values in the logistic regression analysis were not affected by reassigning the KRAS gene amplification cases to the KRAS mutant group (Fig. 5 and Additional file 1: Fig. S2) probably due to the small number of gene amplification cases. The mean specificity and sensitivity between the two segmentations for PVE-and time-corrected SUR MAX in this case are 0.69 and 0.77, respectively (Additional file 1: Table S2), compared to 0.69 and 0.85 for grouping with the wild type (Table 5).
One study has shown association of KRAS gene amplification with lack of response to anti-EGFR therapy in all four gene amplification cases of 53 non-responding cases; however, according to the authors this was not statistically significant and for these cases it was mutually exclusive with KRAS missense mutations [48]. KRAS gene amplification was seen in 0.7% of cases with "de novo resistance to anti-EGFR treatment" [49]. However, while KRAS amplification is a recurrent event in CRC and designated as likely oncogenic by the OncoKB knowledge base [50], at present, KRAS amplification per se is not a contraindication for anti-EGFR therapy [51] and its clinical significance remains uncertain.
In the investigated cohort, most of the observed KRAS missense mutations are in codon 12 (n = 8), but codons 13 (n = 4), 61 (n = 2) and 146 (n = 1) were also represented. While for mutations in codon 61 the reduction in response rate to cetuximab plus chemotherapy compared to wild type was found to be significant, that for mutations in codon 146 was not significant for a small number of cases [9]. Also, KRAS G13D mutations were shown to have potential positive effect on survival compared to other KRAS mutations in a retrospective study; however; further scrutiny of this effect is needed [52,53]. Therefore, if the above findings of the effect of KRAS missense mutation location are confirmed, investigations of the effect of mutation location on 2-[18F]FDG uptake with larger patient cohorts harboring such mutations would be justified.
The ROC and logistic regression curves (Figs. 4 and 5) show that some of the 2-[18F]FDG uptake-based metrics presented have potential to predict KRAS missense mutations in CLM. From the investigated metrics, PVEand uptake time-corrected SUR MAX results in the highest AUC, specificity and sensitivity and therefore seems most promising for mutational status characterization. Making such a prediction based on previous or on intraprocedural PET images can prompt the interventional radiologist to biopsy lesions which may harbor mutations potentially different from those previously established or among different lesions. In this way, early predictions of CLM KRAS mutations, especially for cases with multiple lesions, would allow optimization of immediate interventions [11,14] as well as of future treatments [6][7][8][9][10].
To further establish 2-[18F]FDG uptake as a biomarker of KRAS+ status in CLM, it is important to address the limitations of the presented investigation. Several steps in that direction can be taken. One would be to reduce the uncertainty in SUV by requiring residual activity measurement for each patient injection. Another would be including more lesions with specifically known mutational status to further reduce the labeling error and improve statistics. Use of SUR eliminates uncertainties related to injected activity and other technical and patient-related factors. While several studies have shown that using the SUR-based uptake time correction leads to finding better defined correlations by reducing variance in uptake, this work extends the applicability of this approach to larger uptake time differences.

Conclusion
Reducing the number of potentially mislabeled cases by focusing predominantly on lesions with genomic profile obtained from 2-[18F]FDG PET/CT-guided biopsies, applying partial volume effect and uptake time corrections, and using the ratio of maximum tumor to mean blood uptake in PET/CT scans, allowed for a highly statistically significant separation of colorectal adenocarcinoma liver lesions expressing KRAS missense mutations. Future testing with a larger patient cohort and implementation of these algorithms in the clinic may allow to assign probabilities for the KRAS mutation status of such lesions from the PET images. This may lead to further optimization of interventions and therapy for metastatic colorectal adenocarcinoma in the liver.