Recurrence after Resection of Pancreatic Cancer: Can Radiomics Predict Patients at Greatest Risk of Liver Metastasis?

Abstract

Background:

Liver metastasis (LM) after pancreatic ductal adenocarcinoma (PDAC) resection is common but difficult to predict, and has grave prognosis. We combined preoperative clinicopathological variables and quantitative analysis of CT imaging to predict early LM.

Methods:

We retrospectively evaluated patients with PDAC submitted to resection between 2005–2014 and identified clinicopathological variables associated with early LM. We performed liver radiomic analysis on preoperative contrast-enhanced CT scans and developed a logistic regression classifier to predict early LM (<6 months).

Results:

In 688 resected PDAC patients, there were 516 recurrences (75%). The cumulative incidence of LM at 5 years was 41%, and patients who developed LM first (n=194) had the lowest 1-year OS (34%), compared to 322 patients who developed extrahepatic recurrence first (61%). Independent predictors of time to LM included poor tumor differentiation (HR=2.30; P<0.001), large tumor size (HR=1.17 per 2-cm increase; P=0.048), lymphovascular invasion (HR=1.50; P=0.015), and liver Fibrosis-4 score (HR=0.89 per 1-unit increase; P=0.029) on multivariate analysis. A model using radiomic variables that reflect hepatic parenchymal heterogeneity identified patients at risk for early LM with an area under the receiver operating characteristic curve (AUC) of 0.71; the performance of the model was improved by incorporating preoperative clinicopathological variables (tumor size and differentiation status; AUC=0.74, NPV=0.86).

Conclusions:

We confirm the adverse survival impact of early LM after resection of PDAC. We further show that a model using radiomic data from preoperative imaging combined with tumor-related variables has great potential for identifying patients at high risk for LM and may help guide treatment selection.

Graphical Abstract

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is an aggressive malignancy with a dismal outcome due to its propensity to evolve undetected and disseminate early during oncogenesis¹. Nearly half of newly diagnosed PDAC patients present with obvious metastatic disease². Among those with localized resectable disease, the majority recur, often at distant sites, soon after surgery, suggesting micrometastatic disease is present but undetectable on conventional imaging at the time of diagnosis. The liver is a frequent site of early disease recurrence after resection of PDAC (up to 53% of patients)^3–6 and generally portends a rapid clinical decline. This clinical reality has largely fueled the current trend toward increased use of neoadjuvant therapy (NAT) in patients with resectable PDAC. However, the results of several studies have not proven NAT superior to upfront resection followed by adjuvant therapy^7,8. The reasons for this lack of benefit are unclear but may relate to our inability to identify and target NAT to high-risk patients. It is clear that better patient selection strategies are required.

To date, few studies have specifically investigated risk factors for liver metastasis (LM) after resection of PDAC^6,9. We hypothesized clinicopathological variables available prior to operation, including those that may relate to alterations in the liver microenvironment, could potentially correlate with the risk of LM – for example, assessments that reflect the global liver function such as MELD score and the Fibrosis-4 (FIB-4) score, which is now used with increasing frequency for non-invasive quantification of liver fibrosis^10,11. Moreover, CT-based quantitative imaging features (radiomics) may better detect early liver changes (e.g., premetastatic niche and/or occult micrometastases^12–16) and thus distinguish patients at high risk of early LM. Radiomics involves the extraction and analysis of high dimensional quantitative features from radiographic images that may reflect underlying biology. Specifically, texture analysis of preoperative contrast-enhanced CT, a technique that quantifies global and spatial heterogeneity in pixel intensity values¹⁷ can be used to predict outcomes similar to other biomarkers^17–19.

We hypothesized that CT-based quantitative imaging features related to contrast extravasation, parenchymal permeation and retention in the liver would correlate with the presence of the pre-metastatic niche; thus they would differ between PDAC patients who develop LM and those that do not, and can be used to identify such high risk patients. Herein, we report a retrospective analysis of LM in a large cohort of patients with resected PDAC, and the development of radiomics-based models to predict risk of early LM. Reliable identification of such patients would potentially provide a rational basis for recommending neoadjuvant therapy before proceeding with resection.

Methods

Patients

After approval by the Institutional Review Board (IRB) at Memorial Sloan Kettering Cancer Center (MSKCC), we queried a prospectively maintained database of patients who underwent pancreatic resection with curative intent from 1/1/2005–12/31/2014 and identified all patients with a confirmed diagnosis of PDAC. Exclusion criteria are shown in Figure 1. There were no patients with R2 resections in the cohort. To validate the proposed predictive models, we selected an independent cohort of 29 patients who met similar criteria and were enrolled in an IRB-approved prospective biomarker discovery trial that took place between 2015 and 2018 at MSKCC.

Figure 1.

Schematic representation of patient cohort and recurrence patterns. EH, extra-hepatic; NED, no evidence of disease recurrence.

Clinical and pathologic variables

Demographic, clinical, laboratory, and pathological data were collected from the electronic medical record. The variables included age at resection, gender, body mass index (BMI), presence of diabetes or thyroid disease, bilirubin, margin status, tumor size and differentiation, N stage, lymphovascular invasion (LVI), and perineural invasion (PNI). Race was not included due to a high amount of missing data in the early years. Laboratory values obtained closest to the time of resection were used. FIB-4 score was calculated using the formula $\left[\text{Age}\left(\text{years}\right)\times \text{AST}\left(\text{U}/\text{L}\right)\right]/\left[\text{Platelet count}\left({10}^3/\mu \text{L}\right)\times \sqrt{\text{ALT}}\left(\text{U}/\text{L}\right)\right]$, as described previously¹¹. Biliary obstruction (BO) was defined as either total bilirubin greater than the upper normal limit (1.0 mg/dL at our institution) or history of biliary drainage (BD) procedure for the relief of preoperative jaundice. Tumor differentiation reported as ‘moderate-poor’ or ‘well-moderate’ was recorded as poor or moderate, respectively. Further, LVI and PNI that were reported as “suspected” were classified as “present.”

Quantitative Imaging Features/Radiomic Features

We identified patients who had contrast-enhanced CT scans with portal venous phase performed up to 3 months before resection, and acquired following the administration of 150 mL iodinated contrast (Omnipaque 300, GE Healthcare, Princeton, NJ) at 4.0 mL/s on a multidetector CT scanner (Lightspeed 16 and VCT, GE Healthcare, Madison, WI) with the following parameters: pitch/table speed 0.984–1.375/39.37–27.50 mm; autoMA 220–380; noise index 12.5–14; rotation time 0.7–0.8 ms; scan delay 80–85 s. We excluded patients who had biliary stents or other implants that caused significant imaging artifacts according to the study radiologist. Axial slices reconstructed at 2.5-mm intervals were used to segment the hepatic parenchyma and the major hepatic vessels in a semi-automatic fashion using Scout Liver software (Pathfinder Technologies Inc., Nashville, TN), as described previously²⁰. Incidental benign tumors (e.g., hemangiomas) were excluded from the segmented parenchyma.

Two hundred and fifty-four previously described radiomic features, reflecting heterogeneity in enhancement patterns, were extracted from the segmented liver (Supporting Table 1) using our inhouse software^18,19. Each feature was extracted from all slices of individual CT scans and averaged to obtain a single measure for each patient, using MATLAB R2015a (MathWorks, Natick, MA, USA).

Study Design

Patient follow-up was derived from available imaging reports and accompanying clinical assessments. Recurrence was either proven by biopsy of suspicious lesions or assumed based on emergence of new, highly suspicious imaging findings associated with new symptoms and/or progression over time. The date of first recurrence was defined as the date a suspicious lesion was first noted, even if it was not definitive at the time and was later interpreted as recurrence or metastatic disease. Time to first recurrence was defined as the time from the date of surgery to date of first recurrence. Similarly, the date of LM was defined as the date a liver lesion was first noted, and the time to LM was defined as the time from the date of surgery to date of LM. Patients with no evidence of recurrence who were lost to follow-up were censored at their last imaging time point. Patients with no postoperative imaging studies available were censored. Patients were classified according to the presence and timing of LM or extrahepatic (EH) recurrence and were subdivided into 3 mutually exclusive groups: LM first (with or without synchronous extrahepatic recurrence; LM-F), EH first (without any LM; EH-F; without any LM), or no evidence of recurrent disease (NED).

Statistical analysis

Overall survival (OS) was calculated from the date of surgery to date of death or last follow-up and estimated using Kaplan-Meier methods. Cumulative incidence of first recurrence with specific patterns (EH-F; LM-F) and time to LM were estimated using competing risks methods. Clinicopathological factors associated with time to LM were examined using Fine and Gray competing risks regression methods²¹. Patients who had a recurrence elsewhere but not liver (n=246) or died without any evidence of disease recurrence (n=82) were treated as competing events. A multivariable competing risk regression model was constructed by including factors that were significantly associated with time to LM. In order to assess the impact of different recurrence patterns on survival, we calculated OS from the date of first recurrence until date of death or last follow-up and compared between different recurrence pattern groups (EH-F, LM-F) using the log-rank test.

All statistical analyses were performed using SAS Version 9.3 (SAS Institute, INC., Cary, NC, USA) or R Version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria) using ‘cmprsk’ package. All p-values were two-sided, and p-values <0.05 were considered statistically significant.

Prediction model development

The primary outcome was early LM, defined as LM within 6 months of surgery. Three prediction models were developed using: (1) clinical variables only; (2) quantitative imaging features (radiomics) only; and (3) a combination of clinical and radiomic variables. The patient cohort from 2005–2014 was used for training and designing the prediction model. An independent set of 29 patients, treated between 2015–2018, was used for validation (Supporting Table 2). Patients with early LM were compared with the rest of the cohort. The first model was developed with logistic regression using significant clinical variables (p<0.05) that can be obtained preoperatively. For the radiomic model, the significant features were initially selected using Wilcoxon rank-sum test, after removing the correlated features. A forward selection method was then applied to select the final set of imaging features to design the radiomic-based model with logistic regression classifier (MATLAB 2015b, Mathworks Inc, Natick, MA). Finally, selected imaging and clinical variables were used for the design of a combined clinical-imaging prediction model with logistic regression classifier. The performance of the models was evaluated with area under the receiver operating characteristics curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

Results

Patient characteristics

A total of 921 patients underwent resection of pathologically confirmed PDAC from 2005–2014. We excluded 140 (15.2%) patients who had received NAT, to avoid potential confounding effects of NAT on liver texture. In the final cohort of 688 patients (Table 1), the majority received adjuvant therapy (535; 77.8%), with most receiving gemcitabine-based chemotherapy with or without other agents. Some patients also received standard radiation therapy with concurrent infusional 5-fluorouracil (151; 21.9%).

Table 1.

Cohort characteristics (n=688)

Variable	No. of patients (unless noted)	(%)
Sex, male	351	(51.0)
Age, years – Median [range]	69.5 [35.0–91.1]
BMI, kg/m 2
<20	29	(4.2)
20–25	239	(34.7)
>25–30	261	(37.9)
>30	156	(22.7)
N/A	3	(0.4)
Hypercholesterolemia	293	(42.6)
N/A	1	(0.1)
Diabetes mellitus	193	(28.1)
N/A	1	(0.1)
Thyroid disease	87	(12.6)
N/A	15	(2.2)
FIB-4 score – Median [range]	1.52 [0.07–19.8]
N/A	3	(0.4)
MELD score – Median [range]	7 [0–27]
N/A	3	(0.4)
Neutrophil-to-lymphocyte ratio – Median [range]	3.17 [0.67–19.25]
Lymphocyte-to-monocyte ratio – Median [range]	4 [0.64–31]
N/A	1	(0.1)
Monocyte-to-lymphocyte ratio – Median [range]	0.08 [0–0.32]
AST:ALT ratio – Median [range]	0.81 [0.02–9.12]
Total bilirubin, mg/dL – Median [range]	1.1 [0.1–35.8]
Biliary obstruction *	455	(66.1)
Preoperative biliary drainage
No	352	(51.2)
Yes, transduodenal (ERCP)	312	(45.3)
Yes, percutaneous (PBD)	22	(3.2)
Yes, other (e.g. T-tube)	1	(0.1)
N/A	1	(0.1)
Type of procedure
Pancreaticoduodenectomy (PD)	332	(48.3)
Pylorus-preserving PD (PPPD)	209	(30.4)
Distal pancreatectomy (DP)	135	(19.6)
Total pancreatectomy (TP)	8	(12)
Central pancreatectomy	3	(0.4)
Combined PD + DP	1	(0.1)
Resection margin status
R0	518	(75.3)
R1	170	(24.7)
Differentiation
Well	11	(16)
Moderate	468	(68.0)
Poor	206	(29.9)
Undifferentiated	2	(0.3)
N/A	1	(0.1)
Tumor size, cm – Median [range]	3 [0.3–13]
N/A	9	(13)
T stage (AJCC 8th ed.)
1 (≤2cm)	101	(14.7)
2 (>2 and <4cm)	447	(65.0)
3 (>4cm)	132	(19.2)
N/A	8	(12)
Positive lymph nodes – Median [range]	2 [0–33]
Lymph node ratio – Median [range]	0.11 [0–1]
N stage (AJCC 8th ed.)
0 (negative lymph nodes)	183	(26.6)
1 (1–3 positive lymph nodes)	284	(41.3)
2 (>3 positive lymph nodes)	221	(32.1)
Lymphovascular invasion	505	(73.4)
N/A	6	(0.9)
Perineural invasion	650	(94.5)
Adjuvant therapy
Chemotherapy	535	(77.8)
N/A	44	(6.4)
Chemo-radiation	151	(21.9)
N/A	50	(7.3)

N/A: not applicable (patients with missing data)

^*defined as elevated preoperative total bilirubin and/or preoperative biliary drainage procedure

Recurrence patterns

We first examined recurrence patterns after resection. The median follow-up among survivors was 4.0 (range 0.1–11.3) years. Nine patients (1.3%) had no postoperative imaging and were censored on the first postoperative day. Over the duration of follow-up, 516 patients had a recurrence, whereas 82 died without evidence of disease on their last follow-up (Figure 1). Local recurrence (i.e., recurrence within the remnant pancreas or the resection bed²²) was observed initially in 243 patients and was the most common site of first recurrence, followed by liver (194), lymph nodes (109), lung (100), peritoneum (89), and bone (8). Rare sites of initial recurrence included left supraclavicular (Virchow’s) lymph node (1), lower parasternal (internal mammary) lymph node (1), peri-umbilical (Sister Mary Joseph) nodule (1), and right scrotum (1).

Among 194 patients who developed LM as the first site of recurrence (LM-F), 85 had synchronous EH disease (Figure 1). On the other hand, 322 patients developed initial recurrence at extra-hepatic site(s) (EH-F), of which 241 never developed LM over the duration of follow-up, whereas 81 patients eventually developed LM, with a median time interval of 8 months (range 1.1–68) for a single EH site (n=57) or 4.3 (range 0.9–20.2) months for multiple simultaneous EH sites (n=24) from resection.

The cumulative incidence of any LM at 1, 2, and 5 years post-resection was 27.6% (95%CI: 24.3–31.3%), 36.4% (95%CI: 32.8–40.1%), and 41.1% (95%CI: 37.3–44.8%), respectively. The risk of LM steadily rose for the first 2 years and nearly plateaued after the second year, with a median time to LM of 9.0 (95%CI: 7.9–10.0) months (Figure 2a). The steepest initial increase was noted for LM-F, with a median time to recurrence of 6.5 (95%CI: 5.5–7.3) months, and 88 patients developing LM-F within 6 months post-resection (Figure 2b). In contrast, the incidence of EH-F recurrence continued beyond 2 years, albeit at a slower rate, with a median time to recurrence of 12.3 (95%CI: 11.0–13.6) months.

Figure 2.

(a) Cumulative incidence of liver metastasis (LM). (b) Cumulative incidence of the 3 main recurrence patterns. (c) Overall survival of entire cohort. (d,) Overall survival (OS) after first recurrence, stratified by recurrence patterns. EH-F, extrahepatic first; LM-F, liver metastasis first; NED, no evidence of disease recurrence.

Survival

At the end of the study period, there were a total of 522 deaths, with a median OS of 25 (95%CI: 24–28) months and 5-year OS of 21% (Figure 2c). Importantly, once initial recurrence was documented, the survival of patients who developed LM-F was significantly shorter (median OS: 8.6 [95%CI: 7.3–9.7] months) vs. patients who had EH-F recurrence (median OS: 16.3 [95%CI: 14.0–18.7] months) (Figure 2d, e), suggesting early LM reflects the most aggressive form of PDAC progression.

Risk factors for liver metastasis

We next analyzed prognostic factors for LM, including variables relating to the primary tumor and surgical procedure, and variables that either reflect liver function (e.g., liver function tests and related indices) or could potentially induce derangements in the liver microenvironment (e.g., obesity/BMI, diabetes mellitus, hypercholesterolemia, thyroid disease, biliary obstruction, biliary drainage) (Table 2). On univariate analysis, none of the clinical factors examined were significantly associated with time to any LM. Neutrophil-to-lymphocyte ratio (NLR) and lymphocyte-to-monocyte ratio (LMR), previously shown to correlate with OS^23,24, were not associated with time to LM, whereas monocyte-to-neutrophil ratio (MNR) showed a potential trend towards negative association (HR 0.09 [95%CI: 0.01–2.44]; P=0.15). Of the biochemical parameters examined, FIB-4 score (which is based on age and laboratory values) was negatively associated with time to LM (HR=0.91 [95%CI: 0.84–0.99] per 1-unit increase; P=0.037). Several pathological variables were also associated with time to LM, including poor tumor differentiation (HR=2.40 [95%CI: 1.87–3.06]; P<0.01), tumor size (HR=1.16 [95%CI: 1.004–1.35] per 2-cm increase; P=0.044), positive N stage (HR=1.56 [95%CI: 1.14–2.14] for N1; 1.83 [95%CI: 1.32–2.54] for N2; P<0.001), and LVI (HR=1.81 [95%CI: 1.34–2.44]; P<0.01). No significant association was found between risk of LM and margin status, lymph node ratio, or PNI.

Table 2.

Factors associated with recurrence in the liver

		Univariate			Multivariate
Variable	HR Comparison	HR	95%CI	P-value	HR	95%CI	P-value
PREOPERATIVE CLINICAL FACTORS
BMI
<20	<20 vs. 20–25	0.86	(0.45–1.66)	0.65
25–30	25–30 vs. 20–25	1.14	(0.87–1.51)	0.35
>30	>30 vs. 20–25	1.14	(0.83–1.56)	0.42
Hypercholesterolemia	Present vs. Absent	1.05	(0.72–1.54)	0.80
Diabetes	Yes vs. No	1.01	(0.99–1.02)	0.25
Thyroid disease	Yes vs. No	0.85	(0.59–1.23)	0.39
PREOPERATIVE LABORATORY FACTORS
Neutrophil-to-lymphocyte ratio (NLR)	Per 1-unit increase	1.01	(0.95–1.06)	0.83
Lymphocyte-to-monocyte ratio (LMR)	Per 1-unit increase	0.98	(0.93–1.02)	0.32
Monocyte-to-neutrophil ratio (MNR)	Per 1-unit increase	0.09	(0–2.44)	0.15
AST/ALT ratio	Per 1-unit increase	0.81	(0.52–1.25)	0.34
Total bilirubin	Per 1-unit increase	1.01	(0.98–1.03)	0.54
FIB-4 score	Per 1-unit increase	0.91	(0.84–0.99)	0.037	0.89	(0.82–0.99)	0.029
MELD score	Per 1-unit increase	1.01	(0.99–1.03)	0.37
Location* / Bilary Obstruction (BO)** / Biliary Drainage (BD)***
Proximal / No BO / No BD	Compared to Distal / No BO / No BD	1.20	(0.78–1.86)	0.41	1.21	(0.77–1.91)	0.416
Proximal / BO + / No BD	Compared to Distal / No BO / No BD	1.40	(0.95–2.21)	0.088	1.53	(0.97–2.42)	0.066
Proximal / BO + / BD +	Compared to Distal / No BO / No BD	1.50	(1.03–2.09)	0.033	1.38	(0.94–2.03)	0.101
PATHOLOGICAL FACTORS
Margin status	Positive vs. Negative	0.84	(0.63–1.12)	0.24
Tumor differentiation	Poor/Undifferentiated vs. Well/Moderate	2.40	(1.87–3.06)	<0.01	2.30	(1.78–2.98)	<.001
Tumor Size (cm)	Per 2-cm increase	1.16	(1.00–1.35)	0.044	1.17	(1.00–3.56)	0.048
N Stage
N1	N1 vs. N0	1.56	(1.14–2.14)	<.001
N2	N2 vs. N0	1.83	(1.32–2.54)	<.001
Lymph node ratio	>0.2 vs. ≤0.2	1.15	(0.89–1.49)	0.28
Lymphovascular invasion	Yes vs. No	1.81	(1.34–2.44)	<.01	1.50	(1.08–2.09)	0.015
Perineural invasion	Yes vs. No	1.36	(0.77–2.39)	0.29

^*Location was deducted from the procedure type as either distal (patients who underwent distal pancreatectomy, DP) or proximal (any procedure other than DP, including PD/PPPD, CP, and TP). One patient who had a combined PD+DP was excluded.

^**Biliary obstruction (BO) was defined as elevated preoperative total bilirubin (>1 mg/dL) and/or preoperative biliary drainage (BD) procedure. Four patients with distal tumors and borderline elevated total bilirubin (that would have been classified as having BO despite no clinical or other biochemical evidence) were excluded.

^***One patient with unknown BD procedure was excluded.

Tumor location, BO, and BD interacted as co-variates; therefore, patients were classified into 4 subgroups for analysis: (a) distal tumors (i.e., pancreatic body and tail) without BO or BD, (b) proximal tumors (i.e., pancreatic head) without BO or BD, (c) proximal tumors with BO but not BD, and (d) proximal tumors with BO and BD (Figure 3). Compared to patients with distal tumors, patients with proximal tumors and BO had increased risk of LM, which was statistically significant in the subset of patients who underwent preoperative BD (HR=1.5 [95%CI: 1.03–2.09]; p=0.033) and borderline significant for patients who did not have BD (HR=1.4 [95%CI: 0.95–2.21]; p=0.088). Further, the time to LM did not differ significantly between the 3 subgroups of patients with proximal tumors (with or without BO and BD) (Table 2).

Figure 3.

(a) Patients with tumors of the body/tail of the pancreas (distal tumors) underwent distal pancreatectomy (n=135), whereas those with tumors of the uncinate process, head or neck of the pancreas (proximal tumors) underwent pancreaticoduodenectomy, central, or total pancreatectomy (n=553). The latter were classified based on the presence of biliary obstruction (BO) and biliary drainage (BD) procedures in 3 subgroups. (b) The 4 subgroups were compared for time to liver metastasis. Results of univariate analysis are presented as hazard ratios [95% confidence interval] for each column group (bold) compared to each row group.

On multivariable analysis, FIB-4 remained significantly associated with time to LM in an inverse manner (HR=0.89 [95%CI: 0.82–0.99] per 1-unit increase; p=0.029). Tumor location and BO/BD were not significantly associated with time to LM, although there was a borderline association for proximal tumors with BO (Table 2). Lastly, poor tumor differentiation (HR=2.30 [95%CI: 1.78–2.98]; P<0.001), increasing tumor size (HR=1.17 [95%CI: 1.00–3.56] per 2-cm increase; P=0.048), and LVI (HR=1.50 [95%CI: 1.08–2.09]; P=0.015) remained significantly associated with time to any LM.

Preoperative Prediction of Liver Metastasis

Given the adverse impact of early LM on outcome, we next aimed to design a prediction model targeting this patient subset. We identified 126 patients with suitable preoperative contrast-enhanced CT scans (see Methods) and adequate follow-up, and built clinical and radiomic models to classify the patients into 2 categories: patients who developed LM within 6 months (early LM), and patients who had no evidence of LM at 6 months.

The clinical prediction model, based on tumor size and differentiation, features that can be obtained preoperatively, yielded an AUC of 0.74 and 0.69 for training and validation data, respectively. FIB-4 score was examined but did not improve the predictive accuracy of the model and was therefore excluded. The imaging model, based on 6 significantly associated texture features (IH2; FD49, FD53; LBP78, LBP79, and LBP126), achieved a high specificity (training: 0.92 [95%CI: 0.85–0.97]; validation: 0.86 [95%CI: 0.64–0.97]), with an AUC of 0.74 in the training and 0.71 in the validation set (Figure 4). Although the imaging model provided better AUC and specificity, the clinical model delivered higher sensitivity on the training set. An integrated model combining the imaging and clinical variables (Figure 4b) improved the performance, providing an AUC of 0.74 for the validation data that tended to be significantly better compared to the either the imaging or the clinical model (log-rank test; p=0.067 and p=0.061, respectively). Most importantly, the integrated model delivered high sensitivity, with a NPV of 0.95 for the training and 0.86 for the validation cohort, thus making it a promising tool for discrimination of patients with a low probability of early LM.

Figure 4.

CT radiomic analysis for discrimination of patients at high risk of early liver metastasis. (a) Digital images acquired using standard pancreas protocol contrast-enhanced CT were imported to the PathFinder software and the liver parenchyma was segmented. Each pixel’s gray level intensity was expressed on a numerical scale, and pixels were compared either altogether or one by one based on their spatial relationship to develop 254 radiomic “features” for each patient. (b) Receiver operating curve (ROC) for detection of patients with early liver metastasis (LM <6 months after resection) using a radiomics-based prediction model with 6 features that differed significantly between patients with early LM (<6 months) vs. without LM at 6 months in the training set and a validation set. Performance of the imaging, clinical (using tumor size and differentiation; one patient was missing tumor differentiation status and was removed from this comparison), and combined imaging and clinical prediction models; sensitivity (Sens), specificity (Spec), positive predictive value (PPV), negative predictive value (NPV), and area under the ROC (AUC) are presented as with the 95% confidence interval.

Discussion

In the present study, we characterized the incidence, pattern, and timing of LM in patients with resected PDAC. Our analysis showed that LM is often seen early after resection of PDAC and is a harbinger of poor survival, compared to initial isolated EH recurrence (median OS after first recurrence: 8.6 vs. 16.3 months, respectively). Therefore, development of early LM appears to be associated with aggressive disease biology that progresses more rapidly, whereas initial EH recurrence is suggestive of a relatively more indolent phenotype. Knowledge of the risk of LM prior to resection can thus have important prognostic implications and guide the therapeutic approach such that patients with impending LM can be treated with neoadjuvant chemotherapy in order to spare them from potential operative morbidity related to an operation that would offer little benefit.

To enrich our cohort with patients nearing disease progression (e.g. those harboring subclinical micrometastases), we defined “early LM” as LM within 6 months from resection, and sought predictors of this outcome among several perioperative parameters. Our rationale for selection of these variables was two-fold: first, we included parameters that reflect liver physiology, as this may contribute to the ability of disseminated tumor cells (DTCs) to seed the liver and, conversely, may be altered by DTCs and primary tumor-derived secreted factors, such as cytokines and extracellular vesicles^{12,13,16,25–28}. Second, we examined variables available in the preoperative setting, which could be useful for determining LM risk prior to resection. Among the parameters known preoperatively, FIB-4 score was significantly associated with reduced risk of LM, suggesting that derangements in liver physiology may contribute to conditions unfavorable for DTC engraftment and subsequent LM. This finding is in line with prior observations that cirrhotic patients may have a lower tendency to develop LM from various malignancies^29,30, though abnormal FIB-4 score does not necessarily imply cirrhosis, especially in the context of biliary obstruction. Several primary tumor-related variables were significantly correlated with early LM, most notably poor tumor differentiation (HR=2.30; P<0.001), a finding that has also been observed in prior studies^6,9. Increasing tumor size and LVI also remained significantly associated with increased risk of LM on multivariate analysis. These findings are not unexpected, since poor differentiation and greater tumor size may reflect higher intratumoral heterogeneity and thus a higher likelihood of cancer cell clones able to adapt to the adverse microenvironment of distant organs, while at the same time being more difficult to treat due to emergence of resistance³¹. Prior studies have examined associations between perioperative variables and recurrence-related outcomes and found similar associations of tumor size and differentiation, but not LVI^6,9.

Unlike previous reports, our study is the first to specifically examine the potential association of BO with LM. Biliary obstruction, resulting from tumor encroachment on the distal bile duct, complicates two-thirds of localized PDAC cases and can induce profound effects on liver physiology. Besides alterations of hepatic metabolism, BO leads to changes in extracellular matrix and modulates the hepatic immune microenvironment^32,33. Despite this, BO did not impact the odds of LM among patients with proximal tumors (who are prone to malignant BO) in our analysis. Further, FIB-4 score did not differ between patients who had BO and those who did not (data not shown). Therefore, BO itself does not appear to correlate with LM, although this is not definitive.

Our ultimate goal was to generate a prediction algorithm to preoperatively identify patients at risk of early LM. We hypothesized that radiomic analysis of preoperative liver contrast-enhanced CT would uncover subtle changes that correlate with liver premetastatic niche and/or micrometastatic disease. This was based on preclinical data supporting that the livers of PDAC patients who develop LM exhibit alterations that precede metastasis development, characterized by activation of resident immune cells and other non-parenchymal cells, inflammatory cell recruitment, stromal remodeling, and altered vascular permeability (a state collectively termed the premetastatic niche)^{12,13,15,16,25–28}.

We observed that texture features could predict early LM. Notably, IH2 (i.e., standard deviation), which represents heterogeneity of the liver parenchyma, was higher among patients with early LM, and therefore may be an indicator of occult micrometastatic disease or altered hepatic microenvironment that favors metastatic seeding. Using forward selection, 6 features were combined in a prediction model with high specificity for early LM. The performance of the imaging model was improved by including tumor size and differentiation status, both of which can be obtained preoperatively. Notably, the combined imaging/clinical model had a high sensitivity rather than specificity, making it more suitable in identifying patients with a low likelihood of early LM, who can therefore derive the highest benefit from upfront pancreatic resection. Therefore, we envision identification of patients at very high risk of early LM may be feasible by radiomic analysis of the premetastatic liver at the time of operative planning, and further improved by knowledge of primary tumor differentiation via preoperative core needle biopsy. Future studies to prospectively assess this approach are warranted. Furthermore, radiomic analysis of the pancreas may yield correlative information about the biology of the primary tumor that can supplement or even substitute preoperative needle biopsy, which could improve specificity by identifying more aggressive tumors.

The present study has several limitations. First, this was a retrospective study, and as a result, preoperative CT scans and laboratory tests did not occur at a standardized time point before resection. Second, although our institutional practice is to obtain postoperative follow-up imaging at pre-specified intervals, several patients had follow-up imaging at different time points, either because of postoperative complications/adjustment of treatment plans or because of adjuvant oncologic treatment at other institutions. Third, in the majority of cases, metastatic lesions were not biopsy-proven, although there was no doubt clinically. Also, many patients were excluded from the radiomic analysis, mainly because the imaging parameters deviated from the specified protocol or the study was suboptimal – development of a platform that can standardize CT images acquired on different scanners within a range of settings will enable expanded applicability of our approach. Additionally, we did not include CA19–9 in our analysis because it was available only in 447/688 (65.0%) of the patients. While a subset analysis including CA19–9 suggested that elevated levels may be associated with time to LM, CA19–9 levels did not differ between patients that developed liver metastases and those that did not nor among patients in the NED, EH-F and LM-F subgroups (Supporting Figure 1). Finally, tumor size and differentiation were derived from the final pathology report rather than from preoperative sources; however, both are obtainable preoperatively (e.g., using endoscopic needle biopsy).

In summary, in this proof-of-principle study, we show the potential of liver radiomics of preoperative CT scans to identify high risk of early LM in patients with resectable PDAC. This approach is worthy of further prospective evaluation and may be further enhanced when combined with other “omic” data that reflect tumor biology and propensity for early dissemination, and potentially novel imaging techniques that provide additional information relating to the liver parenchyma, such as magnetic resonance or ultrasound elastography. Further research and optimization of radiomics will be required before it can be used in clinical practice.

Synopsis.

We describe recurrence patterns after pancreatic cancer resection and their impact on survival. We present a model combining preoperative clinicopathological and radiomic variables that predicts risk for early liver metastasis, which may be useful for treatment recommendations.