Sage Journals: Discover world-class research

Abstract

BACKGROUND:

Patients with advanced pancreatic cancer (APC) and liver metastases have much poorer prognoses than patients with other metastatic patterns.

OBJECTIVE:

This study aimed to develop and validate a radiomics model to discriminate patients with pancreatic cancer and liver metastases from those with other metastatic patterns.

METHODS:

We evaluated 77 patients who had APC and performed texture analysis on the region of interest. 58 patients and 19 patients were allocated randomly into the training and validation cohorts with almost the same proportion of liver metastases. An independentsamples $t$ -test was used for feature selection in the training cohort. Random forest classifier was used to construct models based on these features and a radiomics signature (RS) was derived. A nomogram was constructed based on RS and CA19-9, and was validated with calibration plot and decision curve. The prognostic value of RS was evaluated by Kaplan-Meier methods.

RESULTS:

The constructed nomogram demonstrated good discrimination in the training (AUC $=$ 0.93) and validation (AUC $=$ 0.81) cohorts. In both cohorts, patients with RS $>$ 0.61 had much poorer overall survival than patients with RS $<$ 0.61.

CONCLUSIONS:

This study presents a radiomics nomogram incorporating RS and CA19-9 to discriminate patients who have APC with liver metastases from patients with other metastatic patterns.

Keywords

Liver metastases metastatic patterns pancreatic cancer radiomics nomogram texture features

1. Introduction

Pancreatic cancer is one of the most highly lethal diseases with almost as many deaths as cases worldwide [1, 2]. Among all types of distant metastases in APC, liver metastases is the most common, and it occurs in more than half of patients with APC [3, 4]. In our previous study, patients with APC and liver metastases showed much poorer prognoses compared with patients who had other patterns of metastases (6.1 months vs. 10.9 months, $P<$ 0.001) [5]. The discrimination of patients who have APC with liver metastases from patients with other metastatic patterns via texture analysis (TA) can help us understand more about the characteristics of pancreatic cancer and its association with metastatic patterns, which may lead to novel treatment strategies and better prognoses for patients.

Computed tomography (CT) is a commonly used examination for the initial diagnosis and staging of pancreatic cancer. However, the information acquired from CT in the clinical practice is limited due to inadequate elucidation of CT images [6]. For example, the size-based and shape-based features derived from CT images (e.g., longest tumor diameter) are sometimes associated with marked variability in outcomes and response, such as pseudoprogression in patients receiving immunotherapy [7]. Recently, the rapidly evolving field of radiomics has made it possible to translate medical images into high-throughput quantitative data [8, 9], and TA provides an opportunity for more precise diagnosis and treatment [10, 11, 12]. In addition, much evidence has shown that TA may be used to characterize and identify specific subsets of cancers. For example, previous studies have shown that a CT-based radiomics signature (RS) could help identify patients with lung adenocarcinoma patients who have high risks of distant metastases and specific histologic subtypes [13, 14].

In pancreatic cancer, previous studies have shown that radiomics features have both diagnostic and prognostic value. For example, Guo et al. showed that CT imaging features and texture parameters could help differentiate pancreatic neuroendocrine carcinoma from pancreatic ductal adenocarcinoma [15]. Sandrasegaran et al. showed that CT texture features, especially kurtosis and the mean value of positive pixels, were significantly associated with overall survival (OS) in pancreatic cancer [16]. In addition, Attiyeh et al. found that CT texture features were associated with SMAD4 status and stromal content in pancreatic cancer [17]. However, few studies have used radiomics features to characterize specific subtypes of pancreatic cancer.

Therefore, the aim of this study was to develop and validate a radiomics nomogram that incorporated both RS and clinicopathologic variables to discriminate patients who have APC with liver metastases from patients with other metastatic patterns.

2. Patients and methods

2.1 Patients

The primary cohort of this study consisted of 77 patients with APC enrolled in Changzhou No. 2 People’s Hospital. The following inclusion criteria were applied: (1) newly diagnosed and pathologically confirmed pancreatic adenocarcinoma; (2) no concurrent cancer at another organ site; and (3) complete records of baseline clinicopathologic features available. Of all the patients, 40 had liver metastases, and 37 had other metastatic patterns. The patients were randomly allocated into two stratified cohorts that included almost the same proportion of liver metastases: a training cohort ( $n=$ 58) and a validation cohort ( $n=$ 19) (Supplemental Fig. S1). Only the training cohort was used for the initial selection of predictive features. Demographic and clinicopathologic features were collected from the electronic medical records. Informed consent was obtained from each patient and ethical approval was obtained by the ethics committees of Changzhou No. 2 People’s Hospital.

2.2 CT image acquisition

Contrast-enhanced CT examinations were performed on a 128-row dual-source CT scanner (SOMATOM Definition Flash, Siemens, Germany) with 120 kV, tube current modulation and a 1 mm reconstructed section thickness. All patients were instructed to fast for at least 8 hours before administration of 70 ml of intravenous contrast (Iohexol 300 mg/mL) at a rate of 3 mL/s. After injection of the contrast agent, patients underwent double helical scanning at the arterial phase and portal venous phase while holding their breaths. The portal venous phase was selected as a region of interest (ROI), because pancreatic cancer boundaries have been most consistently seen in this phase [17, 18].

2.3 Image processing

ROIs were drawn on each slice with primary pancreatic cancer and metastatic tumors on the portal venous phase using the Semantic Segmentation Editor (Hitachi Automotive and Industry Laboratory). Then all ROIs were imported into a freeware software (Local Image Features Extraction, or LIFEx, version 5.10, https://www.lifexsoft.org/) for TA [19]. In the segmented tumors, the volume of interest (VOI) and the histogram were calculated as first-order features [20]. To calculate second- and higher-order texture features, the number of grey levels used to resample the ROI content was set to 64.0, as described in previous studies [20, 21]. Spatial resampling was 2.0 mm (X-direction), 2.0 mm (Y-direction), and 1.0 mm (Z-direction) in Cartesian coordinates. Texture features were evaluated by four texture matrices, including the grey-level co-occurrence matrix (GLCM), the grey-level run length matrix (GLRLM), the neighborhood grey-level different matrix (NGLDM) and the grey-level zone length matrix (GLZLM). In total, 36 features were extracted from the TA (Fig. 1). The values of features from the largest cross section of each ROI were used for additional analysis.

Figure 1.

Workflow of image processing and model construction. Abbreviations: RFC, Random Forest Classifier; RS, radiomics signature.

Table 1

Baseline clinicopathologic characteristics of patients with APC

Characteristics	Training cohort ( $n=$ 58)		Validation cohort ( $n=$ 19)		$P$ -value
Age (median, SD)	66.5	(10.9)	68	(10.1)	0.814
Gender
Male	33	(56.9%)	11	(57.9%)	0.939
Female	25	(43.1%)	8	(42.1%)
Body mass index (kg*m ${}^{-2}$ )
Underweight $<$ 18.5	14	(24.1%)	3	(15.8%)	0.807
Normal weight (18.5–25)	38	(65.5%)	14	(73.7%)
Overweight (25–30)	5	(8.6%)	2	(10.5%)
Obese $>$ 30	1	(1.7%)	0	(0.0%)
Hypertension
Yes	20	(34.5%)	3	(15.8%)	0.122
No	38	(65.5%)	16	(84.2%)
Diabetes mellitus
Yes	24	(41.4%)	10	(52.6%)	0.391
No	34	(58.6%)	9	(47.4%)
Smoking
Yes	10	(17.2%)	3	(15.8%)	0.883
No	48	(82.8%)	16	(84.2%)
Drinking
Yes	4	(6.9%)	1	(5.3%)	0.802
No	54	(93.1%)	18	(94.7%)
Liver metastases
Yes	30	(51.7%)	10	(52.6%)	0.945
No	28	(48.3%)	9	(47.4%)
Primary tumor location
Head and neck	28	(48.3%)	8	(42.1%)	0.640
Body and tail	30	(51.7%)	11	(57.9%)
CA19-9 (U/ml)
$<$ 1000	36	(62.1%)	13	(68.4%)	0.617
$\geqslant$ 1000	22	(37.9%)	6	(31.6%)

Table 2

Comparison between baseline clinicopathologic characteristics of patients with APC according to liver metastases

Characteristics	Training cohort					Validation cohort
	With liver metastases ( $n=$ 30)		Without liver metastases ( $n=$ 28)		$P$ -value	With liver metastases ( $n=$ 10)		Without liver metastases ( $n=$ 9)		$P$ -value
Age (median, SD)	67	(11.2)	65	(10.7)	0.571	67	(11.7)	67	(8.7)	0.951
Gender
Male	17	(56.7%)	16	(57.1%)	0.971	6	(60.0%)	5	(55.6%)	0.845
Female	13	(43.3%)	12	(42.9%)		4	(40.0%)	4	(44.4%)
Body mass index (kg*m ${}^{2}$ )
Normal weight (18.5–25)	19	(63.3%)	19	(67.9%)	0.717	7	(70.0%)	7	(77.8%)	0.701
Abnormal weight	11	(36.7%)	9	(32.1%)		3	(30.0%)	2	(22.2%)
Hypertension
Yes	7	(23.3%)	13	(46.4%)	0.064	3	(30.0%)	0		0.211
No	23	(76.7%)	15	(53.6%)		7	(70.0%)	9	(100.0%)
Diabetes mellitus
Yes	11	(36.7%)	13	(46.4%)	0.451	5	(50.0%)	5	(55.6%)	0.809
No	19	(63.3%)	15	(53.6%)		5	(50.0%)	4	(44.4%)
Smoking
Yes	6	(20.0%)	4	(14.3%)	0.732	3	(30.0%)	0		0.211
No	24	(80.0%)	24	(85.7%)		7	(70.0%)	9	(100.0%)
Drinking
Yes	3	(10%)	1	(3.6%)	0.612	1	(10.0%)	0		0.330
No	27	(90.0%)	27	(96.4%)		9	(90.0%)	9	(100.0%)
Primary tumor location
Head and neck	13	(43.3%)	15	(53.6%)	0.436	2	(20%)	6	(66.7%)	0.070
Body and tail	17	(56.7%)	13	(46.4%)		8	(80%)	3	(33.3%)
CA19-9 (U/ml)
$<$ 1000	15	(50%)	21	(75.0%)	0.050	5	(50.0%)	8	(88.9%)	0.141
$>$ 1000	15	(50%)	7	(25%)		5	(50.0%)	1	(11.1%)
Radiomics signature	0.68	(0.13)	0.38	(0.17)	$<$ 0.001	0.60	(0.19)	0.47	(0.13)	0.102

2.4 Statistical analysis

Statistical analysis was conducted with R software (version 3.6.1, Institute for Statistics and Mathematics, Vienna, Austria) and Python (version 3.6.8). The univariable correlations between clinicopathologic characteristics and liver metastases were assessed with the Pearson Chi-Square test or Fisher’s exact test for categoric variables and with an independent-samples $t$ -test for continuous variables. An independentsamples $t$ -test was used to investigate the differences in texture features in patients with APC and liver metastases versus patients with other metastatic patterns. The correlations between texture parameters were assessed by Pearson’s correlation coefficient with the R package “psych”. Random Forest Classifier (RFC) was used with the Python packages (scikit-learn, version 0.21.0, https://scikit-learn.org/). The receiver operating characteristic (ROC) curves were drawn by Python package Matplotlib (https://matplotlib.org/). The RS was derived from the model, which combined classifiers by averaging their probabilistic predictions, and the tool X-tile was used to determine the optimal cutoff value of the RS [22]. A nomogram was constructed according to the RS and the CA19-9 level with the R package “rms”. Internal validation of the nomogram was performed with the calibration plot using bootstrapping with 1000 re-samples. The Kaplan-Meier analysis and the log-rank test were performed to show the stratification of OS in each group at different levels of the RS.

3. Results

3.1 Patient characteristics

Table 1 demonstrated the baseline clinicopathologic characteristics of patients with APC in both the training cohort and the validation cohort. All the variables, including age, gender, body mass index (BMI), hypertension, diabetes mellitus, smoking, drinking, liver metastases, primary tumor location and CA19-9 were comparable between the two groups. In the training cohort, 30 patients (51%) had liver metastases, and 28 patients (48.3%) had other metastatic patterns. In the validation cohort, 10 patients (52.6%) had liver metastases and 9 patients (47.4%) had other metastatic patterns.

3.2 Correlations between clinicopathologic characteristics and liver metastases

In the training cohort, patients with liver metastases had a higher level of CA19-9 compared with patients without liver metastases ( $P=$ 0.050). However, other factors including age, gender, BMI, hypertension, diabetes mellitus, smoking, drinking, and primary tumor location were comparable between the two groups. In the validation cohort, one (11.1%) in nine patients without liver metastases had a CA19-9 level $>$ 1000 whereas five (50%) in 10 patients with liver metastases had a CA19-9 level $>$ 1000 ( $P=$ 0.141, Table 2).

Table 3
Principal parameters calculated by texture analysis in the training cohort

Texture parameter	With liver metastases ( $n=$ 30)				Without liver metastases ( $n=$ 28)				$P$ -value
	Mean		SD		Mean		SD
VOI volume (ml)	2.	757	1.	485	2.	390	1.	407	0.	334
Histogram
Skewness	$-$ 0.	449	0.	669	$-$ 0.	138	0.	712	0.	095 ${}^{*}$
Kurtosis	5.	216	1.	503	4.	475	2.	044	0.	128
Entropy	30.	560	4.	730	33.	814	6.	091	0.	034 ${}^{**}$
Energy	0.	0417	0.	0071	0.	0386	0.	0110	0.	227
GLCM
Homogeneity	0.	272	0.	037	0.	255	0.	04	0.	097 ${}^{*}$
Energy	0.	00373	0.	00125	0.	0368	0.	00202	0.	904
Contrast	59.	49	23.	958	72.	033	33.	637	0.	114
Correlation	0.	579	0.	125	0.	607	0.	105	0.	360
Entropy	388.	57	129.	37	414.	36	154.	49	0.	485
Dissimilarity	5.	651	1.	247	6.	288	1.	532	0.	093 ${}^{*}$
GLRLM
SRE	0.	951	0.	011	0.	957	0.	012	0.	062 ${}^{*}$
LRE	1.	224	0.	56	1.	194	0.	063	0.	069 ${}^{*}$
LGRE	0.	00408	0.	00215	0.	00541	0.	00305	0.	070 ${}^{*}$
HGRE	1562.	272	462.	718	1318.	077	438.	232	0.	049 ${}^{**}$
SRLGE	0.	00399	0.	00206	0.	00531	0.	00303	0.	067 ${}^{*}$
SRHGE	1484.	772	445.	664	1261.	858	420.	662	0.	060 ${}^{*}$
LRLGE	0.	00448	0.	00260	0.	00585	0.	00321	0.	089 ${}^{*}$
LRHGE	1919.	981	553.	156	1575.	718	525.	744	0.	024 ${}^{**}$
GLNU	26.	250	16.	124	21.	051	13.	719	0.	195
RLNU	559.	140	287.	426	497.	449	286.	597	0.	410
RP	0.	935	0.	0146	0.	943	0.	0160	0.	061 ${}^{*}$
NGLDM
Coarseness	0.	0155	0.	0101	0.	0151	0.	0093	0.	894
Contrast	0.	308	0.	196	0.	312	0.	220	0.	943
Busyness	0.	0440	0.	0237	0.	0391	0.	0224	0.	416
GLZLM
SZE	0.	830	0.	037	0.	832	0.	082	0.	936
LZE	2.	307	0.	444	2.	098	0.	618	0.	155
LGZE	0.	00450	0.	00213	0.	00589	0.	00327	0.	070 ${}^{*}$
HGZE	1547.	275	456.	013	1283.	562	424.	670	0.	032 ${}^{**}$
SZLGE	0.	00410	0.	00208	0.	0546	0.	00332	0.	077 ${}^{*}$
SZHGE	1290.	796	391.	341	1089.	213	357.	343	0.	050 ${}^{**}$
LZLGE	0.	00662	0.	00438	0.	00823	0.	00444	0.	181
LZHGE	3677.	151	1266.	982	2769.	051	1128.	107	0.	010 ${}^{**}$
GLNU	19.	451	10.	990	15.	979	10.	158	0.	218
ZLNU	328.	217	154.	711	305.	244	174.	888	0.	591
ZP	0.	769	0.	049	0.	822	0.	188	0.	156

Figure 2.

The correlations between texture parameters.

3.3 Correlation between texture parameters and liver metastases

In the training cohort, no significant differences in VOI and NGLDM parameters were observed between the two groups with independentsamples $t$ -tests ( $P>$ 0.1, Table 3). In the histogram analysis, patients without liver metastases, versus those with metastases showed greater values of skewness ( $P=$ 0.092) and entropy ( $P=$ 0.026). In addition, in the GLCM analysis, only homogeneity ( $P=$ 0.092) and dissimilarity ( $P=$ 0.087) differed between patients with liver metastases versus those with other metastatic patterns. In the GLRLM analysis, patients with liver metastases had lower values of SRE ( $P=$ 0.054), LGRE ( $P=$ 0.059), SRLGE ( $P=$ 0.056), LRLGE ( $P=$ 0.078) and RP ( $P=$ 0.053), but greater values of LRE ( $P=$ 0.061), HGRE ( $P=$ 0.044), SRHGE ( $P=$ 0.056), and LRHGE ( $P=$ 0.019) compared with those who had other metastatic patterns. Furthermore, in the GLZLM analysis, patients with liver metastases, compared with those with other metastases, had greater levels of HGZE ( $P=$ 0.027), SZHGE ( $P=$ 0.046) and LZHGE ( $P=$ 0.006), but patients with other metastatic patterns compared with those with liver metastases, had greater levels of LGZE ( $P=$ 0.058) and SZLGE ( $P=$ 0.065).

Pearson’s correlation coefficient was used to investigate the correlations between texture parameters in the training cohort (Fig. 2). The results demonstrated significant correlations between some pairs of these parameters. For example, skewness showed significant negative correlations with GLRLM_HGRE ( $r=-$ 0.84), GLRLM_SRHGE ( $r=-$ 0.84), GLRLM_LRHGE ( $r=-$ 0.84), GLZLM_HGZE ( $r=-$ 0.82) and GLZLM_SZHGE ( $r=-$ 0.81). Conversely, GLZLM_ZLNU showed significant positive correlations with VOI ( $r=$ 0.96), GLRLM_GLNU ( $r=$ 0.83), GLRLM_RLNU ( $r=$ 0.98) and GLZLM_GLNU ( $r=$ 0.89).

Figure 3.

The ROC curve and AUC of Random Forest Classifier models in the training cohort.

Figure 4.

Developed radiomics nomogram to discriminate advanced pancreatic cancer patients with liver metastases or other metastatic patterns.

Figure 5.

Calibration plot and decision curve of the radiomics nomogram.

Figure 6.

Kaplan-Meier estimates OS according to radiomics signature in the training cohort (A) and validation cohort (B).

3.4 Development and validation of the radiomics nomogram

Eighteen texture parameters (shown in Table 3) with significance values $<0.1$ were used as inclusion criteria for additional model construction with RFC. The model demonstrated an acceptable discrimination ability with RFC, with an area under the curve (AUC) of 0.92 in the training cohort and an AUC of 0.77 in the validation cohort (Fig. 3, Supplemental Fig. S2). An RS was derived from the RFC model, which combined classifiers by averaging their probabilistic predictions.

Given the significant difference in the CA19-9 level between patients with or without liver metastases, a nomogram containing both the RS and CA19-9 was constructed to predict liver metastases (Fig. 4). The nomogram showed good discrimination with an AUC of 0.93 in the training cohort and an AUC of 0.81 in the validation cohort. Internal validation of the nomogram was performed with a calibration plot (Fig. 5A). The decision curve analysis of the RS and the nomogram is presented in Fig. 5B. The decision curve showed that, if the threshold probability was $>$ 85%, the nomogram with both RS and CA19-9 would add more benefit than the nomogram with RS alone.

3.5 Prognostic value of the RS

We also investigated the prognostic value of the RS with the Kaplan-Meier methods and the log-rank test. In the training cohort, patients with an RS $>$ 0.61 had poorer OS than patients with an RS < 0.61 (3.5 months vs 8.0 months, $P<$ 0.001, Fig. 6A). Similar results were observed in the validation cohort (Fig. 6B). Patients with an RS $>$ 0.61 also had much poorer OS versus patients with an RS $<$ 0.61 (3.7 months vs. 9.2 months, $P=$ 0.003).

4. Discussion

In this study, we showed a strong discrimination ability between patients who had APC with liver metastases and patients with other metastatic patterns by developing an RS. A nomogram was constructed using both the RS and CA19-9, and it showed good performance in the training cohort (AUC $=$ 0.93) and in the validation cohort (AUC $=$ 0.81). In addition, the RS also successfully stratified patients according to OS.

Previous studies have shown that a variety of texture features correlates with metastasis in cancer. For example, kurtosis and GLCM-Energy have been associated with progression-free survival and tumor metastasis-dependent factors, such as yes-associated protein [23, 24]. Beckers et al. [25] found that uniformity was an independent and significant predictor of liver metastases in patients with colorectal cancer. Sheen et al. [21] demonstrated that GLZLM_SZLGE was an independent predictor of metastasis in osteosarcoma. Becker et al. [26] also found that texture features from the histogram and the gray-level matrices could be used to predict metastases in mice. In this study, we found that 18 texture features correlated with liver metastases; among them, GLZLM_LZHGE showed the smallest $P$ value of 0.006. The positive correlation of GLZLM_LZHGE indicates that many long homogeneous zones with high grey levels dominate in the fine-grained texture of the tumor. However, whether such phenomenon is a result of tumor cellular metabolism must be investigated in future studies.

Conventional multiparametric analysis has been widely used in previous studies for parameter selection and model construction [12]. Recently, machine-learning algorithms have shown superiority in this aspect, especially in medical imaging analysis [27, 28, 29, 30, 31]. To construct radiomics models in this study, 18 potential predictors were selected from 36 candidate radiomics features. This study used RFC to acquire an RS, which showed an acceptable discrimination ability (AUC $=$ 0.92 in the training cohort and AUC $=$ 0.77 in the validation cohort). This result may reflect the random feature selection and the appropriate randomness to provide accuracy without overfitting [32].

The CA19-9 level did not demonstrate enough discrimination ability alone on the basis of univariable association with liver metastases ( $P=$ 0.050 in the training cohort and $P=$ 0.141 in the validation cohort), which made it seem unnecessary for inclusion in the model. However, in another study of us about the metastatic patterns of pancreatic cancer that included a much larger sample size, we found that the CA19-9 level significantly correlated with liver metastases ( $P=$ 0.001) [33]. Other studies also have found that CA19-9 is an independent risk factor of liver metastases in pancreatic cancer [34]. CA19-9 is usually considered a measure of pancreatic tumor burden, and it is reasonable that patients with larger tumor burdens would be more likely to develop liver metastases. Thus, the lack of significance of CA19-9 in the robustness of the results in this study may be due to the relatively small sample size and confounding by other predictors [35]. We opted to include CA19-9 in the nomogram, and the decision curve showed that if the threshold probability was $>$ 85%, the nomogram with both RS and CA19-9 would offer more benefit than one with RS alone. In addition, the AUC of the nomogram would increase from 0.77 for RS alone to 0.81 for the combination in the validation cohort.

This study had several limitations. First, the retrospective design and relatively small sample size can result in unavoidable bias and reduce. Second, the texture features of the liver were not considered. Recently, Lee [4] found that hepatocytes can direct the formation of a pro-metastatic niche in the liver by coordinating myeloid cell accumulation and fibrosis, which suggests that liver characteristics may affect the metastasis of pancreatic cancer. Previous studies have shown that a TA of a whole-liver CT may predict which patients with colorectal cancer are at risk of developing early ( $<$ 6 months) liver metastases [25]. Third, other factors, such as the presence of hepatitis B virus infection that may increase the risk of liver metastases, were not included in this study [36]. Additional analyses with large sample sizes are needed to address these limitations.

In conclusion, this study presents a radiomics nomogram incorporating both the RS and CA19-9. This nomogram can be used to discriminate patients who have APC with liver metastases from patients with APC and other metastatic patterns.

Authors’ contributions

Conception: T.Z., J.H. and L.W.

Interpretation or analysis of data: T.Z., X.D., Y.Z., J.H. and L.W.

Preparation of the manuscript: T.Z. and J.H.

Revision for important intellectual content: X.D. and M.L.

Supervision: J.H. and L.W.

Funding

This research was funded by grants from the National Natural Science Foundation of China (81902955), Natural Science Foundation of Jiangsu Province (BK201901 61), Project of Jiangsu Shuangchuang Doctor (QT2019 04), Foundation of Changzhou Sci & Tech Program (CJ20190096), the Youth Science and Technology Project of Changzhou Health and Family Planning Commission (QN201817).

Supplementary data

The supplementary files are available to download from http://dx.doi.org/10.3233/CBM-210190.

Footnotes

Acknowledgments

None.

Conflict of interest

The authors declare no conflict of interest.

References

Bray

Ferlay

Soerjomataram

Siegel

R.L.

Torre

L.A.

and Jemal

, Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA Cancer J Clin 68 (2018), 394–424.

Kamisawa

Wood

L.D.

Itoi

and Takaori

, Pancreatic cancer, Lancet 388 (2016), 73–85.

Disibio

and French

S.W.

, Metastatic patterns of cancers: results from a large autopsy study, Arch Pathol Lab Med 132 (2008), 931–939.

Lee

J.W.

Stone

M.L.

Porrett

P.M.

Thomas

S.K.

Komar

C.A.

J.H.

Delman

Graham

Gladney

W.L.

Hua

Black

T.A.

Chien

A.L.

Majmundar

K.S.

Thompson

J.C.

Yee

S.S.

O’Hara

M.H.

Aggarwal

Xin

Shaked

Gao

Liu

Borad

M.J.

Ramanathan

R.K.

Carpenter

E.L.

de Beer

M.C.

de Beer

F.C.

Webb

N.R.

and Beatty

G.L.

, Hepatocytes direct the formation of a pro-metastatic niche in the liver, Nature 567 (2019), 249–252.

Hang

Zhu

Sun

Wang

Pan

Zheng

and Jiang

, Prediction of overall survival for metastatic pancreatic cancer: Development and validation of a prognostic nomogram with data from open clinical trial and real-world study, Cancer Med 7 (2018), 2974–2984.

Chu

L.C.

Goggins

M.G.

and Fishman

E.K.

, Diagnosis and Detection of Pancreatic Cancer, Cancer J 23 (2017), 333–342.

Chiou

V.L.

and Burotto

, Pseudoprogression and Immune-Related Response in Solid Tumors, J Clin Oncol 33 (2015), 3541–3543.

Lambin

Leijenaar

R.T.H.

Deist

T.M.

Peerlings

de Jong

E.E.C.

van Timmeren

Sanduleanu

Larue

Even

A.J.G.

Jochems

van Wijk

Woodruff

van Soest

Lustberg

Roelofs

van Elmpt

Dekker

Mottaghy

F.M.

Wildberger

J.E.

and Walsh

, Radiomics: the bridge between medical imaging and personalized medicine, Nat Rev Clin Oncol 14 (2017), 749–762.

Gillies

R.J.

Kinahan

P.E.

and Hricak

, Radiomics: Images Are More than Pictures, They Are Data, Radiology 278 (2016), 563–577.

10.

Chicklore

Goh

Siddique

Roy

Marsden

P.K.

and Cook

G.J.

, Quantifying tumour heterogeneity in 18F-FDG PET/CT imaging by texture analysis, Eur J Nucl Med Mol Imaging 40 (2013), 133–140.

11.

Holli-Helenius

Salminen

Rinta-Kiikka

Koskivuo

Bré»žk

Bostré°‰

and Parkkola

, MRI texture analysis in differentiating luminal A and luminal B breast cancer molecular subtypes - a feasibility study, BMC Med Imaging 17 (2017), 69.

12.

Huang

Y.Q.

Liang

C.H.

Tian

Liang

C.S.

Chen

Z.L.

and Liu

Z.Y.

, Development and Validation of a Radiomics Nomogram for Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer, J Clin Oncol 34 (2016), 2157–2164.

13.

Parmar

Grossmann

Quackenbush

Lambin

Bussink

Mak

and Aerts

H.J.

, Exploratory Study to Identify Radiomics Classifiers for Lung Cancer Histology, Front Oncol 6 (2016), 71.

14.

Coroller

T.P.

Grossmann

Hou

Rios Velazquez

Leijenaar

R.T.

Hermann

Lambin

Haibe-Kains

Mak

R.H.

and Aerts

H.J.

, CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma, Radiother Oncol 114 (2015), 345–350.

15.

Guo

Zhuge

Wang

Xiao

Wang

Feng

and Chen

, The differentiation of pancreatic neuroendocrine carcinoma from pancreatic ductal adenocarcinoma: the values of CT imaging features and texture analysis, Cancer Imaging 18 (2018), 37.

16.

Sandrasegaran

Lin

Asare-Sawiri

Taiyini

and Tann

, CT texture analysis of pancreatic cancer, Eur Radiol 29 (2019), 1067–1073.

17.

Attiyeh

M.A.

Chakraborty

McIntyre

C.A.

Kappagantula

Chou

Askan

Seier

Gonen

Basturk

Balachandran

V.P.

Kingham

T.P.

D’Angelica

M.I.

Drebin

J.A.

Jarnagin

W.R.

Allen

P.J.

Iacobuzio-Donahue

C.A.

Simpson

A.L.

and Do

R.K.

, CT radiomics associations with genotype and stromal content in pancreatic ductal adenocarcinoma, Abdom Radiol (NY) 44 (2019), 3148–3157.

18.

Eilaghi

Baig

Zhang

Karanicolas

Gallinger

Khalvati

and Haider

M.A.

, CT texture features are associated with overall survival in pancreatic ductal adenocarcinoma - a quantitative analysis, BMC Med Imaging 17 (2017), 38.

19.

Nioche

Orlhac

Boughdad

, Reuz? Goya-Outi

Robert

Pellot-Barakat

Soussan

Frouin

and Buvat

, LIFEx: A Freeware for Radiomic Feature Calculation in Multimodality Imaging to Accelerate Advances in the Characterization of Tumor Heterogeneity, Cancer Res 78 (2018), 4786–4789.

20.

Mungai

Verrone

G.B.

Pietragalla

Berti

Addeo

Desideri

Bonasera

and Miele

, CT assessment of tumor heterogeneity and the potential for the prediction of human papillomavirus status in oropharyngeal squamous cell carcinoma, Radiol Med 124 (2019), 804–811.

21.

Sheen

Kim

Byun

B.H.

Kong

C.B.

Song

W.S.

Cho

W.H.

Lim

S.M.

and Woo

S.K.

, Metastasis risk prediction model in osteosarcoma using metabolic imaging phenotypes: A multivariable radiomics model, PLoS One 14 (2019), e0225242.

22.

Camp

R.L.

Dolled-Filhart

and Rimm

D.L.

, X-tile: a new bio-informatics tool for biomarker assessment and outcome-based cut-point optimization, Clin Cancer Res 10 (2004), 7252–7259.

23.

Lee

C.K.

Jeong

S.H.

Jang

Bae

Kim

Y.H.

Park

Kim

S.K.

and Koh

G.Y.

, Tumor metastasis to lymph nodes requires YAP-dependent metabolic adaptation, Science 363 (2019), 644–649.

24.

Ytre-Hauge

Dybvik

J.A.

Lundervold

, Salvesen ? Krakstad

Fasmer

K.E.

Werner

H.M.

Ganeshan

Hé´Œvik

Bjé´•ge

Trovik

and Haldorsen

I.S.

, Preoperative tumor texture analysis on MRI predicts high-risk disease and reduced survival in endometrial cancer, J Magn Reson Imaging 48 (2018), 1637–1647.

25.

Beckers

R.C.J.

Lambregts

D.M.J.

Schnerr

R.S.

Maas

Rao

S.X.

Kessels

A.G.H.

Thywissen

Beets

G.L.

Trebeschi

Houwers

J.B.

Dejong

C.H.

Verhoef

and Beets-Tan

R.G.H.

, Whole liver CT texture analysis to predict the development of colorectal liver metastases-A multicentre study, Eur J Radiol 92 (2017), 64–71.

26.

Becker

A.S.

Schneider

M.A.

Wurnig

M.C.

Wagner

Clavien

P.A.

and Boss

, Radiomics of liver MRI predict metastases in mice, Eur Radiol Exp 2 (2018), 11.

27.

Fujima

Shimizu

Yoshida

Kano

Mizumachi

Homma

Yasuda

Onimaru

Sakai

Kudo

and Shirato

, Machine-Learning-Based Prediction of Treatment Outcomes Using MR Imaging-Derived Quantitative Tumor Information in Patients with Sinonasal Squamous Cell Carcinomas: A Preliminary Study, Cancers (Basel) 11 (2019).

28.

Zhang

Ouyang

Dong

Zhang

Huang

Tian

and Zhang

, Radiomic machine-learning classifiers for prognostic biomarkers of advanced nasopharyngeal carcinoma, Cancer Lett 403 (2017), 21–27.

29.

Parmar

Grossmann

Rietveld

Rietbergen

M.M.

Lambin

and Aerts

H.J.

, Radiomic Machine-Learning Classifiers for Prognostic Biomarkers of Head and Neck Cancer, Front Oncol 5 (2015), 272.

30.

Pesapane

Codari

and Sardanelli

, Artificial intelligence in medical imaging: threat or opportunity? Radiologists again at the forefront of innovation in medicine, Eur Radiol Exp 2 (2018), 35.

31.

W.L.

Hosny

Schabath

M.B.

Giger

M.L.

Birkbak

N.J.

Mehrtash

Allison

Arnaout

Abbosh

Dunn

I.F.

Mak

R.H.

Tamimi

R.M.

Tempany

C.M.

Swanton

Hoffmann

Schwartz

L.H.

Gillies

R.J.

Huang

R.Y.

and Aerts

, Artificial intelligence in cancer imaging: Clinical challenges and applications, CA Cancer J Clin 69 (2019), 127–157.

32.

Breiman

, Random Forests, Machine Learning 45 (2001), 5–32.

33.

Zhu

Zhou

Zhang

Hang

and Zee

B.C.

, Clinical significance of site-specific metastases in pancreatic cancer: a study based on both clinical trial and real-world data, J Cancer 12 (2021), 1715–1721.

34.

G.Y.

Y.H.

S.X.

M.Z.

Hao

C.Q.

and S

L.Z.

, Risk factors of liver metastasis from advanced pancreatic adenocarcinoma: a large multicenter cohort study, World J Surg Oncol 15 (2017), 120.

35.

Sauerbrei

Royston

and Binder

, Selection of important variables and determination of functional form for continuous predictors in multivariable model building, Stat Med 26 (2007), 5512–5528.

36.

Wei

X.L.

Qiu

M.Z.

Chen

W.W.

Jin

Ren

Wang

Luo

H.Y.

Wang

Z.Q.

Zhang

D.S.

Wang

F.H.

Y.H.

and Xu

R.H.

, The status of HBV infection influences metastatic pattern and survival in Chinese patients with pancreatic cancer, J Transl Med 11 (2013), 249.

Development and validation of a radiomics nomogram to discriminate advanced pancreatic cancer with liver metastases or other metastatic patterns

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSIONS:

Keywords

1. Introduction

2. Patients and methods

2.1 Patients

2.2 CT image acquisition

2.3 Image processing

3. Results

3.1 Patient characteristics

3.2 Correlations between clinicopathologic characteristics and liver metastases

Table 3 Principal parameters calculated by texture analysis in the training cohort

3.5 Prognostic value of the RS

4. Discussion

Authors’ contributions

Funding

Supplementary data

Footnotes

Acknowledgments

Conflict of interest

References

Table 3
Principal parameters calculated by texture analysis in the training cohort