Baseline tumor growth rate highlights the heterogeneity of well differentiated gastroenteropancreatic neuroendocrine tumors and predicts for increases in Ki67 index over time

Refined risk stratification for gastroenteropancreatic neuroendocrine tumors (GEP‐NETs) has the potential to improve comparisons of study populations across clinical trials and facilitate drug development. Tumor growth rate (TGR) is a radiological metric with demonstrated prognostic value in well differentiated grade 1 and 2 (G1–2) GEP‐NETs, but little is known about TGR in G3 NETs. In this retrospective study of 48 patients with advanced G1–3 GEP‐NET, we calculated baseline TGR (TGR0) from radiological images of metastases acquired prior to first‐line therapy and evaluated its association with disease characteristics and outcomes. The median pretreatment Ki67 proliferation index for G1–3 tumors combined was 5% (range = 0.1%–52%) and median TGR0 was 4.8%/month (m) (range = 0%–45.9%/m). TGR0 correlated with pretreatment Ki67 across G1–3 pooled and within G3 GEP‐NET. Patients with higher TGR0 (>11.7%/m) tumors, which were primarily G3 pancreatic NETs, exhibited decreased time to first therapy (median, 2.2 vs. 5.3 months; p = .03) and shorter overall survival (median, 4.1 years vs. not reached; p = .003). Independent of therapies given, higher TGR0 GEP‐NETs experienced a greater incidence of Ki67 increase (100 vs. 50%; p = .02) and greater magnitude of Ki67 change (median, 14.0 vs. 0.1%; p = .04) upon serial biopsy. Importantly, TGR0, but not grade, predicted for future Ki67 increase in this series. Given the heterogeneity of well differentiated GEP‐NETs, future clinical trials may benefit from stratification for TGR0, particularly in G1–2 tumors, in which TGR0 does not correlate with Ki67. TGR0 has the potential to noninvasively identify patients with previously undiagnosed grade progression and those in whom more or less frequent monitoring may be appropriate. Additional research is needed to determine the prognostic and predictive value of TGR0 in larger and more homogeneously treated cohorts, and to ascertain if post‐treatment TGR has value in previously treated patients starting a new line of therapy.

study of 48 patients with advanced G1-3 GEP-NET, we calculated baseline TGR (TGR 0 ) from radiological images of metastases acquired prior to first-line therapy and evaluated its association with disease characteristics and outcomes. The median pretreatment Ki67 proliferation index for G1-3 tumors combined was 5% (range = 0.1%-52%) and median TGR 0 was 4.8%/month (m) (range = 0%-45.9%/m). TGR 0 correlated with pretreatment Ki67 across G1-3 pooled and within G3 GEP-NET. Patients with higher TGR 0 (>11.7%/m) tumors, which were primarily G3 pancreatic NETs, exhibited decreased time to first therapy (median, 2.2 vs. 5.3 months; p = .03) and shorter overall survival (median, 4.1 years vs. not reached; p = .003). Independent of therapies given, higher TGR 0 GEP-NETs experienced a greater incidence of Ki67 increase (100 vs. 50%; p = .02) and greater magnitude of Ki67 change (median, 14.0 vs. 0.1%; p = .04) upon serial biopsy. Importantly, TGR 0 , but not grade, predicted for future Ki67 increase in this series. Given the heterogeneity of well differentiated GEP-NETs, future clinical trials may benefit from stratification for TGR 0 , particularly in G1-2 tumors, in which TGR 0 does not correlate with Ki67. TGR 0 has the potential to noninvasively identify patients with previously undiagnosed grade progression and those in whom more or less frequent monitoring may be appropriate. Additional research is needed to determine the prognostic and predictive value of TGR 0 in larger and more homogeneously treated cohorts, and to ascertain if post-treatment TGR has value in previously treated patients starting a new line of therapy. poorly differentiated neuroendocrine carcinomas (NECs). 2,7,8 Ki67 proliferation index is commonly used to define eligibility criteria for clinical trials or as a variable in the assessment of treatment response in well differentiated NETs. Pivotal trials establishing the antitumor activity of somatostatin analogs were performed in patients with low Ki67 index tumors. Ki67 index <10% was required for the CLARINET study, while >90% of patients in the PROMID study had tumors with Ki67 index ≤2%. 9,10 More recently, the NETTER-1 study of lutetium-177 ( 177 Lu)-Dotatate peptide receptor radionuclide therapy (PRRT) restricted enrollment to patients with progressive mid-gut NETs with Ki67 index ≤20%. 11 Phase III studies evaluating the oral tyrosine kinase inhibitor (TKI) surufatinib in panNETs (SANET-p) and extrapancreatic NETs (SANET-ep) enrolled patients with well differentiated G1-2 NETs (Ki67 ≤ 20%). 12,13 Further, several recent studies assessing immune checkpoint inhibitors in NENs incorporated Ki67 index in their exploratory analyses. [14][15][16] Despite the ongoing use of Ki67 index to stratify patients in trials, assessment of Ki67 requires invasive biopsies and can yield heterogenous results within and across tumors of the same patient and over time. 17 Thus, alternative metrics complementary to Ki67 that can serve as predictors of prognosis and treatment efficacy are needed.
While discordance in patients with multiple tumor samples (e.g. multifocal primary tumors or synchronous metastases) at baseline that are enough to change grade are relatively uncommon, baseline Ki67 index values are typically higher in metastases. 17,18 Furthermore, several reports suggest that grade progression over time is common in patients undergoing serial tumor biopsies (e.g., in up to 83% of GEP-NETs). 17,[19][20][21] Tumor growth rate (TGR) is a radiological metric that describes the percentage change in tumor volume per month. One method by which TGR is calculated uses the Response Evaluation Criteria in Solid Tumors (RECIST) sums of target lesions, along with the time between tumor evaluations. 22,23 A post hoc analysis of the CLARINET trial, as well as the retrospective GREPONET and GREPONET-2 studies, demonstrated the prognostic value of baseline TGR (TGR 0 ) for G1-2 NET, with on-treatment TGR at 3 months (TGR 3m ) emerging as a potentially powerful endpoint as it integrates treatment effects. [24][25][26] In an analysis of G1-2 GEP-NET, TGR 3m but not TGR 0 or grade, was an independent factor for progression in patients receiving therapy. 25 In low grade GEP-NET, TGR 0 did not differ by tumor site or grade, and TGR 3m did not depend on grade or Ki67 index. 26 Importantly, little is known about TGR 0 in G3 NET and its rela-

| TGR calculation
TGR 0 was defined as the percent change in RECIST-measurable metastatic target lesion size per month (%/m) using two pretreatment scans. When possible, the scan immediately prior to the pretreatment tumor biopsy was chosen as the first pretreatment scan.
In order to minimize TGR 0 error, we identified the second pretreatment scan as the scan furthest in time from but no later than one year following the first. For the TGR 0 calculation, TGR 0 = 100 Â (e TG -1), where TG = 3 Â log (D2/D1)/t. 22,23 D1 and D2 represented the sum of the longest diameters (SLD) at dates 1 and 2, respectively, while t = (date 2 À date 1 + 1)/30.44 in months. SLD for target, but not non-target or new, lesions were quantified according to RECIST 1.1 guidelines. 27 A maximum of five total target lesions were considered. If there were more than two lesions per organ, then the two largest upon first scan were chosen unless they could not yield reproducible repeated measurements.
Because 4/48 (8%) first scans and 17/48 (35%) second scans were acquired after primary tumor resection, we elected to exclude measurements of all primary tumors for consistency. For patients with any pre-or post-treatment biopsies subsequent to baseline, longitudinal Ki67 change was determined by subtracting the baseline Ki67 from the largest subsequent Ki67 value. In order to address the predictive value of TGR 0 , longitudinal Ki67 change was only calculated for patients whose subsequent biopsies were obtained after their second pretreatment TGR scan.

| Statistical analysis
Statistical analyses were performed with GraphPad Prism 8.0 and R version 4.0.5. Statistical significance was declared based on p < .05.
No multiple testing adjustment was performed.
For all G1-3 patients pooled, Ki67 ranged from 0.1% to 52% with a median of 5%. Additional patient and disease characteristics are summarized in Table 1.

| TGR 0 characterization
We sought to understand the interpatient heterogeneity of TGR 0 and its relationship with Ki67 indices and tumor grade. TGR 0 measurements were determined for 48 patients with a total of 92 target lesions-73 (79.3%) hepatic and 19 (20.7%) nodal, including nine involving the mesentery after originating from small bowel primaries.

| Association between TGR 0 and serial biopsy results
To better understand the association between TGR 0 and prognosis, we investigated pre-and post-treatment serial biopsy results 17    knowledge, prior work on TGR 0 has been limited to G1-2 NET. [24][25][26] We found that TGR 0 correlates with Ki67 across pooled G1-3 NETs, and this finding is primarily driven by the G3 subgroup. TGR 0 , but not grade, predicted for future Ki67 increase in this series. Furthermore, repeat tumor biopsies following suspicious clinical behavior all showed Ki67 increases ( Figure 4D). Our findings are consistent with previous reports of grade progression and Ki67 index migration in both pancreatic and small bowel NETs. 7,8,20,30 The data suggest that TGR 0 may provide information that is clinically relevant and distinct from grade. An unexpectedly high TGR 0 could alert clinicians to a nonrepresentative baseline biopsy (and the potential value of a repeat biopsy) and therefore prompt repeat tumor assessments sooner rather than later.
Like the GREPONET-2 study, most (79%) of the target lesions used for TGR 0 calculations were in the liver. 25 Similar to the CLARI-NET study, which was limited to tumors with Ki67 index <10%, there was no correlation between TGR 0 and Ki67 in G1 or G2 tumors. 24 We also found no significant correlation between TGR 0 value and the sum of longest tumor diameters (i.e., smaller tumors did not appear to have a different TGR 0 than larger tumors). In terms of target lesion selection (e.g., metastatic site and tumor size), we elected to abide by RECIST criteria in choosing the largest and/or most well-described lesions (minimum long axis of 1 cm for non-nodal lesions, minimum short axis of 1.5 cm for nodal lesions), excluding new and non-target lesions, and using a maximum of two target lesions per organ and five overall. Mesenteric nodal disease can be challenging to assess due to fibrosis. 28,29 However, in our series, the correlation between TGR 0 and pretreatment Ki67 remained even when mesenteric lesions were excluded, suggesting additional work in this area is needed.
Limitations of our study include its retrospective nature and small sample size. Additionally, the requirement for two pretreatment scans at least 1 month apart for eligibility likely biased against inclusion of patients with aggressive tumors. Thus, the true variance of TGR 0 in well differentiated GEP-NETs may be even wider than depicted here (0-46%/m). We also found that site of origin was not equally distributed across grades: small bowel NETs were overrepresented in the G1 NET group, pancreas NETs were overrepresented in G2/3 NET, and colorectal NETs were rare across all grades. This is a potentially important finding, recognizing that site of origin impacts molecular features, tumor biology, and overall survival. [31][32][33] Finally, it is worth recognizing that, when assessing the ability of TGR 0 to predict longitudinal Ki67 change, we considered any non-zero positive change as an increase in Ki67 upon serial biopsy. One could have defined a minimum threshold of change based on either relative or absolute Ki67 increase, but the cutoff value that is clinically relevant has not been established.
Several open questions remain regarding TGR as a biomarker.
First, the optimal interval between imaging studies for TGR analyses remains unclear. Unlike the prospective CLARINET study, images for our retrospective TGR analysis were acquired at variable intervals.
Perhaps not surprisingly, the time between scans correlated negatively with TGR 0 , suggesting higher TGR 0 is associated with aggressive clinical behavior that prompts earlier imaging. Another potential explanation is that the accuracy of a real-time TGR estimate might be reduced when the scan interval is long, such that recent changes in growth rate are underappreciated. Second, the impact of interval treatment on TGR also warrants further study (e.g., second-line studies and beyond). Even first-line therapy with somatostatin analogs (SSAs) decreases TGR at 3 months (TGR 3m ), suggesting TGR 0 is best measured in treatment-naïve patients. [24][25][26] Patients in our cohort received different therapies over time, and it is possible that certain treatments may alter TGR such that the relationship between Ki67 and TGR is no longer preserved. The potential value of on-and post-treatment TGR for assessing tumor response, resistance, and grade progression remains to be fully elucidated. Third, prospective studies are needed to determine the TGR 0 cutoff most applicable to clinical practice. Of note, while our optimal TGR 0 cutoff of 11.7%/m is higher than the cutoff identified in other published reports, these studies have widely varying eligibility criteria, including maximum tumor grade, number of prior therapies, and types of therapy received during the study periods. [24][25][26] In addition, optimal TGR 0 cutoffs differed for pancreatic versus GI tract NETs writingreview and editing. Eric Nakakura: Methodology; writingreview and editing. Spencer C. Behr: Methodology; writingreview and editing. Nancy Joseph: Data curation; investigation; methodology; writingreview and editing. Li Zhang: Formal analysis; methodology; supervision; writingreview and editing.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available deidentified upon reasonable request from the corresponding author.
The data are not publicly available due to privacy or ethical restrictions.