A multidimensional nomogram combining imaging features and clinical factors to predict the invasiveness and metastasis of combined hepatocellular cholangiocarcinoma.

BACKGROUND: Combined hepatocellular cholangiocarcinoma (CHCC-CCA) is a rare type of primary liver cancer having aggressive behavior. Few studies have investigated the prognostic factors of CHCC-CCA. Therefore, this study aimed to establish a nomogram to evaluate the risk of microvascular invasion (MVI) and the presence of satellite nodules and lymph node metastasis (LNM), which are associated with prognosis.
METHODS: One hundred and seventy-one patients pathologically diagnosed with CHCC-CCA were divided into a training set (n=116) and validation set (n=55). Logistic regression analysis was used to assess the relative value of clinical factors associated with the presence of MVI and satellite nodules. The least absolute shrinkage and selection operator (LASSO) algorithm was used to establish the imaging model of all outcomes, and to build clinical model of LNM. Nomograms were constructed by incorporating clinical risk factors and imaging features. The model performance was evaluated on the training and validation sets to determine its discrimination ability, calibration, and clinical utility. Kaplan Meier analysis and time dependent receiver operating characteristic (ROC) were displayed to evaluate the prognosis value of the predicted nomograms of MVI and satellite nodule.
RESULTS: A nomogram comprising the platelet to lymphocyte ratio (PLR), albumin-to-alkaline phosphatase ratio (AAPR) and imaging model was established for the prediction of MVI. Carcinoembryonic antigen (CEA) level and size were combined with the imaging model to establish a nomogram for the prediction of the presence of satellite nodules. Favorable calibration and discrimination were observed in the training and validation sets for the MVI nomogram (C-indexes of 0.857 and 0.795), the nomogram for predicting satellite nodules (C-indexes of 0.919 and 0.883) and the LNM nomogram (C-indexes of 0.872 and 0.666). Decision curve analysis (DCA) further confirmed the clinical utility of the nomograms. The preoperatively predicted MVI and satellite nodules by the combined nomograms achieved satisfactory performance in recurrence-free survival (RFS) and overall survival (OS) prediction.
CONCLUSIONS: The proposed nomograms incorporating clinical risk factors and imaging features achieved satisfactory performance for individualized preoperative predictions of MVI, the presence of satellite nodules, and LNM. The prediction models were demonstrated to be good indicator for predicting the prognosis of CHCC-CCA, facilitating treatment strategy optimization for patients with CHCC-CCA. 2021 Annals of Translational Medicine. All rights reserved.

Introduction

Combined hepatocellular cholangiocarcinoma (CHCC-CCA) is a rare form of primary liver malignancy that exhibits features of both hepatocytic and cholangiocytic differentiation. CHCC-CCA accounts for approximately 2–5% of all primary liver cancer cases, and 5-year survival rate of CHCC-CCA is 17.7% (1,2). Owing to its low incidence rate, many problems regarding its diagnostic, therapeutic, and prognostic features have not yet been thoroughly understood or defined.

Although there have been substantial studies investigating both preoperative clinical or radiologic variables individually or in combination to predict the prognostic factors of HCC preoperatively, such as microvascular invasion (MVI) and lymph node metastasis (LNM) (3-5), the established models targeting HCC are not suitable for CHCC-CCA since the proportions of hepatocytic and cholangiocytic characteristics in the lesion cannot be determined prior to biopsy or surgery. Notably, imaging features can depict a variety of lesion properties. For example, the Liver Imaging Reporting and Data System (LI-RADS) incorporates imaging features into a standardized categorization system for the detection of hepatic lesions (6). The imaging features of CHCC-CCA show considerable variation and can be classified into LR-M or LR-4/LR-5, accounting for 32–63% and 7–72%, respectively, according to recent studies. This system was demonstrated to have both diagnostic and prognostic value, with reported studies suggesting that CHCC-CCA patients categorized as LR-M had a worse prognosis than those categorized as LR-4/LR-5 (7,8).

Moreover, whether the clinical or pathological characteristics of CHCC-CCA resemble HCC or ICC is still controversial (9). Some scholars have reported the presence of hepatitis virus infection and elevated α-fetoprotein (AFP), similar to HCC, while others have reported serum CA19-9 elevation and incomplete capsule formation, similar to ICC (10,11). It was also found that CHCC-CCA had the propensity to mimic HCC with respect to venous invasion and satellite formation, while it resembled ICC with respect to LNM and bile duct involvement (12). Owing to these sophisticated differences, it is crucial to develop a specific model to explore the aggressive biologic behavior associated with the prognosis of CHCC-CCA. To the best of our knowledge, no study has yet combined imaging features with clinical characteristics to predict related prognostic factors. Therefore, the aim of this study was to develop and validate a novel multidimensional nomogram that incorporates both for preoperative prediction of the crucial factors associated with prognosis, including MVI, satellite nodules and LNM, which could facilitate a more individualized treatment approach. The prognostic values of the prediction models were also explored.

We presented the following article in accordance with the TRIPOD reporting checklist (available at https://dx.doi.org/10.21037/atm-21-2500).

Methods

Patient selection and study design

The study flow chart for this study is shown in . We retrospectively screened the clinical records of patients with postoperative pathologically confirmed CHCC-CCA between January 2015 and December 2019, which were retrieved from the information system of Zhongshan Hospital, Fudan University. All of the patients received curative surgery. CHCC-CCA was confirmed according to pathological diagnosis referring to the 2010 WHO classification system for CHCC-CCA. A total of 146 patients were classified as classical type, 18 patients were classified as typical subtype of stem cell features, five patients were classified as intermediate cell subtype of stem cell features and two patients were classified as cholangiolocellular subtype of stem cell features (13). The inclusion criteria were as follows: (I) patients with clear pathologically confirmed CHCC-CCA after liver resection; (II) Child-Pugh A or B liver function; (III) patients with detailed preoperative laboratory data; and (IV) MRI examinations performed within 30 days before hepatectomy of sufficient quality to satisfy the diagnostic criteria. Additionally, patients were excluded if they met the following exclusion criteria: (I) previous treatment history, such as transarterial chemoembolization (TACE) therapy and neoadjuvant chemotherapy, and (II) synchronous or previous malignancies. Subsequently, 171 CHCC-CCA patients were randomly divided into the training set (n=116) and validation set (n=55) using “Caret” R package (14). This retrospective study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Clinical Research Ethics Committee of Zhongshan Hospital of Fudan University (B2019-139R) and individual consent for this retrospective analysis was waived.

Figure 1

Study flowchart. (A) Flowchart detailing the patient selection process and exclusion criteria; (B) analysis workflow of this study. CHCC-CCA, combined hepatocellular cholangiocarcinoma; LASSO, least absolute shrinkage, and selection operator; LNM, lymph node metastasis; MVI, microvascular invasion; ROC, receiver operating characteristic; DCA, Decision curve analysis.

Liver MRI

All patients were scanned with a 1.5-T MR scanner (Magnetom Aera, Siemens Healthcare). Routine liver protocols consisted of transverse T2WI, T1WI in-phase and opposed-phase, and diffusion-weighted imaging (DWI, b value =0, 50, and 500 s/mm2). Dynamic imaging was performed with a T1-weighted fat-suppressed sequence. The arterial phase was acquired when the contrast agent (Gd-DTPA, Magnevist, Bayer HealthCare) reached the ascending aorta after the intravenous administration of Gd-DTPA at a dose of 0.1 mmol/kg at a rate of 2 mL/s. The portal venous phase and delayed phase were operated on at 60–70 and 160–180 s, respectively.

Preoperative clinicopathological and imaging variables

Baseline and clinicopathological variables were collected, including age, sex, etiology, laboratory tests and tumor characteristics (i.e., tumor size, tumor number and tumor location). Receiver operating characteristic (ROC) curves were utilized to determine the optimal cutoff value of the continuous variables for diagnosing MVI, and then, the continuous variables were converted to categorical variables for analysis. Additionally, the following qualitative imaging parameters were evaluated on a plain MR scan: (I) contour; (II) hemorrhage; (III) hepatic capsular retraction; and (IV) peritumoral bile duct dilation. Dynamic enhancement characteristics were as follows: (I) arterial enhancement pattern; (II) washout; (III) peritumoral enhancement pattern, which was defined as the enhancing portion surrounding the tumor border (15); (IV) false capsule; and (V) delayed central enhancement. Other imaging features were as follows: (VI) targetoid diffusion restriction; (VII) corona enhancement; (VIII) nodule in nodule architecture; (IX) mosaic architecture; and (X) diffusion restriction. Reviewers assigned a LI-RADS version 2018 (LI-RADS v2018) category for each hepatic observation in the study population, namely, LR-TIV (tumor in vein), LR-M (definitely or probably malignant, not HCC-specific) or LR 1–5 (1, definitely benign; 2, probably benign; 3, indeterminate probability of HCC; 4, probably HCC; or 5, definitely HCC) (16,17). Two experienced radiologists independently reviewed the MR images. Both radiologists were aware that all patients had a high risk of liver primary cancer but were blinded to the specific diagnosis of each tumor. A consensus was reached after discussion in case of any discrepancies. Only the largest pathologically confirmed CHCC-CCA was evaluated. All surgical specimens were examined histopathologically to identify the presence of MVI, satellite nodules and LNM. MVI was defined as the presence of a tumor in the portal vein, hepatic vein, or large capsular vessel of the surrounding hepatic tissue lined by endothelium that was visible only by microscopy. Satellite nodules were defined as detached micrometastatic nodules of HCC cells embedded in the hepatic parenchyma. The pathological LNM status of each patient was identified according to the histopathological reports.

Development of the combination nomograms

Univariate and multivariate logistic regression analyses were performed to evaluate the independent clinical risk factors for the presence of MVI and satellite nodules in the training set. Considering multicollinearity among the imaging variables, the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm was used to select the most useful predictive imaging features. With regard to the small number of events relative to the number of clinical variables for predicting LNM, LASSO regression was used to select both the clinical and imaging predictors of LNM. Using the coefficients derived from the LASSO logistic regression models, we then constructed a formula to calculate the risk score for each patient. Formula = expression index1 × βindex1 +…+ expression index n × β index n (where β is the regression coefficient derived from LASSO regression) (18,19). Independent risk clinical factors and imaging risk scores from the LASSO regression model were integrated into the nomogram for predicting MVI and satellite nodules. In addition, both the clinical risk scores and imaging risk scores obtained from the LASSO regression models were used to construct a nomogram for predicting LNM. The clinical model and imaging model were established separately using the independent clinical risk factors and the imaging risk scores obtained from the LASSO regression.

Evaluation and validation of the combination nomograms and their prognostic value

The performance of the proposed nomograms was evaluated in terms of three aspects: discrimination ability, calibration, and clinical utility. ROC curves were plotted, followed by Harrell’s C-index (20) and calibration curves, with the goal of validating the discrimination ability and accuracy of the constructed nomograms in both the training and validation sets (21,22). Decision curve analysis (DCA) assigns a weight to the relative value of the benefit to harm, thereby modeling the consequences of clinical decisions (23). The decision curve was plotted at different threshold probabilities for the overall patients. Kaplan-Meier curves and log rank tests were conducted to determine the recurrence-free survival (RFS) and overall survival (OS) differences between different risk groups based on the median risk score, which was calculated by the formulas constructed by the predicted nomograms of the MVI and satellite nodules. Time-dependent ROC curves were displayed to evaluate the performances of the predicted nomograms for the MVI and satellite nodules.

Statistical analysis

All statistical analyses were performed using R software (version 3.5.2; https://www.r-project.org/). Student’s t test was used to compare continuous variables with a normal distribution, and the Mann-Whitney U test was used to analyze variables with an abnormal distribution. Categorical variables were compared using the chi-square test or Fisher’s exact test. The reported statistical significance levels were all two-sided, and P<0.05 was considered significant.

Results

Study flowchart and clinicopathological characteristics

In total, 171 consecutive patients with pathologically confirmed CHCC-CCA after hepatectomy between January 2015 and December 2019 who met the inclusion criteria were included. Then, 116 and 55 patients were assigned to the training set and validation set, respectively ().

The preoperative clinicopathologic and MR imaging characteristics of the patients are listed in and . Representative CHCC-CCA images assigned as different LI-RADS categories are shown in and . There was no significant difference in terms of the clinicopathologic and imaging characteristics between the training and validation sets. Histopathologically identified MVI was observed in 48 (41.4%) and 27 (49.1%) patients in the two sets, respectively, and satellite nodules were observed in 21 (18.1%) and 9 (16.4%) patients in the two sets, whereas LNM was observed in 6 (5.2%) and 5 (9.1%) patients in the two sets, respectively. No statistically significant difference existed in the MVI rate (P=0.342), satellite nodule rate (P=0.780) or LNM rate (P=0.329) between the two sets, verifying their use as training and validation sets.

Table 1

Comparisons of clinical characteristics of patients in the training and validation set

Variable	Training set (N=116) (%)	Validation set (N=55) (%)	P value*
Age, year			0.780
<65	95 (81.9)	46 (83.6)
≥65	21 (18.1)	9 (16.4)
Gender			0.232
Male	90 (77.6)	38 (69.1)
Female	26 (22.4)	17 (30.9)
Histopathological type			0.378
Classical type	108 (93.1)	38 (69.1)
Typical subtype of stem cell features	14 (12.1)	4 (7.3)
Intermediate cell subtype of stem cell features	4 (3.4)	1 (1.8)
Cholangiolocellular subtype of stem cell features	2 (1.7)	0 (0.0)
Hepatitis B virus			0.389
Negative	19 (16.4)	12 (21.8)
Positive	97 (83.6)	43 (78.2)
CA19-9, ng/mL			0.367
<37	87 (60.3)	45 (83.6)
≥37	28 (39.7)	10 (16.4)
AFP, ng/mL			0.781
<400	90 (78.3)	42 (76.4)
≥400	25 (21.7)	13 (23.6)
CEA, ng/mL			0.761
<5	94 (81.7)	46 (83.6)
≥5	21 (18.3)	9 (16.4)
Albumin, g/L			0.229
<35	3 (2.6)	0 (0.0)
≥35	113 (97.4)	55 (100.0)
ALT, U/L			0.931
<40	83 (71.6)	39 (70.9)
≥40	33 (28.4)	16 (29.1)
AST, U/L			0.738
<40	88 (75.9)	43 (78.2)
≥40	28 (24.1)	12 (21.8)
GGT, U/L			0.390
<50	63 (54.3)	26 (47.3)
≥50	53 (45.7)	29 (52.7)
Total bilirubin, μmol/L			0.190
<17.1	89 (76.7)	37 (67.3)
≥17.1	27 (23.3)	18 (32.7)
Direct bilirubin, μmol/L			0.219
<6.8	95 (82.6)	41 (74.5)
≥6.8	20 (17.4)	14 (25.5)
Size, cm			0.914
<5	70 (60.9)	33 (60.0)
≥5	45 (39.1)	22 (40.0)
No. of tumors			0.479
Solitary	94 (81.0)	42 (76.4)
Multiple	22 (19.0)	13 (23.6)
Cirrhosis			0.649
No	29 (25.0)	12 (21.8)
Yes	87 (75.0)	43 (78.2)
NLR			0.205
<2.19	75 (64.7)	30 (54.5)
≥2.19	41 (35.3)	25 (45.5)
PLR			0.173
<76.74	37 (31.9)	12 (21.8)
≥76.74	79 (68.1)	43 (78.2)
LMR			0.244
<3.03	32 (27.6)	20 (36.4)
≥3.03	84 (72.4)	35 (63.6)
AAPR			0.530
<0.53	51 (44.0)	27 (49.1)
≥0.53	65 (56.0)	28 (50.9)
APRI			0.204
<0.16	43 (37.1)	26 (47.3)
≥0.16	73 (62.9)	29 (52.7)
ANRI			0.630
<14.73	89 (76.7)	44 (80.0)
≥14.73	27 (23.3)	11 (20.0)
Satellite nodules			0.780
Absent	95 (81.9)	46 (83.6)
Present	21 (18.1)	9 (16.4)
MVI			0.342
Absent	68 (58.6)	28 (50.9)
Present	48 (41.4)	27 (49.1)
LNM			0.329
Absent	110 (94.8)	50 (90.9)
Present	6 (5.2)	5 (9.1)
Capsule			0.838
None	69 (59.5)	35 (63.6)
Complete	23 (19.8)	9 (16.4)
Incomplete	24 (20.7)	11 (20.0)
Tumor location			0.492
Right liver lobe	29 (25.0)	11 (20.0)
Left liver lobe	84 (72.4)	44 (80.0)
Other location	3 (2.6)	0 (0.0)

Data are shown as number of patients, with the percentage in parentheses. *, P value was calculated by χ2 test or Fisher’s exact test. CA 19-9, carbohydrate antigen 19-9; AFP, α‐fetoprotein; CEA, carcinoembryonic antigen; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyltranspeptidase; NLR, neutrophil to lymphocyte ratio; PLR, platelet to lymphocyte ratio; LMR, lymphocyte to monocyte ratio; AAPR, albumin-to-alkaline phosphatase ratio; APRI, aspartate aminotransferase to platelet ratio index; ANRI, aspartate aminotransferase to neutrophil ratio index; MVI, microvascular invasion; LNM, lymph node metastasis.

Table 2

Comparisons of MR imaging features of patients in the training and validation set

Variable	Training set (N=116) (%)	Validation set (N=55) (%)	P value*
Shape			0.888
Globular	44 (37.9)	23 (41.8)
Lobulated	56 (48.3)	25 (45.5)
Irregular	16 (13.8)	7 (12.7)
Contour			0.888
Non-smooth	73 (62.9)	34 (61.8)
Smooth	43 (37.1)	21 (38.2)
Hemorrhage			0.766
Absent	95 (81.9)	44 (80.0)
Present	21 (18.1)	11 (20.0)
Targetoid restriction			0.413
Absent	89 (76.7)	39 (70.9)
Present	27 (23.3)	16 (29.1)
Restricted diffusion			0.764
Absent	10 (8.6)	4 (7.3)
Present	106 (91.4)	51 (92.7)
Arterial enhancement pattern			0.099
None	0 (0.0)	2 (3.6)
Global	67 (57.8)	28 (50.9)
Peripheral	49 (42.2)	25 (45.5)
Peritumoral enhancement pattern			0.413
None	71 (61.2)	28 (50.9)
Wedge	22 (19.0)	12 (21.8)
Irregular	23 (19.8)	15 (27.3)
Washout			0.711
Absent	37 (31.9)	16 (29.1)
Present	79 (68.1)	39 (70.9)
Delayed central enhancement			0.507
Absent	83 (71.6)	42 (76.4)
Present	33 (28.4)	13 (23.6)
Corona enhancement			0.466
Absent	98 (84.5)	44 (80.0)
Present	18 (15.5)	11 (20.0)
False capsule			0.191
None	56 (48.3)	19 (34.5)
Complete	21 (18.1)	15 (27.3)
Incomplete	39 (33.6)	21 (38.2)
Bile duct dilation			0.165
Absent	70 (60.3)	27 (49.1)
Present	46 (39.7)	28 (50.9)
Surface retraction			0.755
Absent	89 (76.7)	41 (74.5)
Present	27 (23.3)	14 (25.5)
Nodule in nodule architecture			0.633
Absent	105 (90.5)	51 (92.7)
Present	11 (9.5)	4 (7.3)
Mosaic architecture			0.244
Absent	70 (60.3)	28 (50.9)
Present	46 (39.7)	27 (49.1)
LI-RADS categorization			0.526
LR-3	8 (6.9)	3 (5.5)
LR-M	32 (27.6)	21 (38.2)
LR-4/LR-5	53 (45.7)	20 (36.4)
LR-TIV	23 (19.8)	11 (20.0)

*, P value was calculated by χ2 test or Fisher’s exact test. MR, magnetic resonance; LI-RADS, Liver Imaging Reporting and Data System.

Figure 2

Images of a 66-year-old male patient with CHCC-CCA. There was an irregular tumor in left lobe of liver showing mild hyperintensity on T2-weighted imaging (A) and hyperintensity diffusion-weighted imaging (B). T1-weighted imaging (C) showed heterogeneous hypointensity. It presented non-rim arterial phase hyperenhancement (D), nonperipheral washout appearance on portal venous phase (E), and enhancing capsule (arrow) on delayed phase (F). CHCC-CCA, combined hepatocellular cholangiocarcinoma.

Figure 3

Images of a 47-year-old male patient with CHCC-CCA. There is a ground tumor (diameter: 2.0 cm) in right lobe of liver showing heterogeneous hyperintensity on T2-weighted imaging (A). It presents hyperintensity on diffusion-weighted imaging (B) and hypointensity on ADC (C). T1-weighted imaging (D) shows homogeneous hypointensity of the lesion. Rim arterial phase hyperenhancement on contrast-enhanced T1-weighted imaging (E) and peripheral washout appearance on portal venous phase (F) is presented on this lesion. CHCC-CCA, combined hepatocellular cholangiocarcinoma; ADC, apparent diffusion coefficient.

Independent prognostic factors of MVI, satellite nodules and LNM and nomogram construction

Univariate analysis revealed that neutrophil to lymphocyte ratio (NLR) (P=0.049), platelet to lymphocyte ratio (PLR) (P=0.001), lymphocyte to monocyte ratio (LMR) (P=0.047), albumin-to-alkaline phosphatase ratio (AAPR) (P=0.008), tumor size (P=0.006) and AFP level (P=0.01) were associated with the presence of MVI, and tumor size (P<0.001) and carcinoembryonic antigen (CEA) (P=0.045) were associated with the presence of satellite nodules. In the multivariate analysis, PLR (OR, 4.8; 95% CI: 1.1–19.7; P=0.031), AAPR (OR, 4.8; 95% CI: 1.3–17.3; P=0.017) were identified as independent predictors of MVI (). Additionally, tumor size (OR, 4.3; 95% CI: 1.1–16.1; P=0.003) and CEA (OR, 12.3; 95% CI: 1.4–107.1; P=0.023) were significantly associated with the presence of satellite nodules (). We utilized the LASSO logistic regression algorithm to select candidate imaging features that were reliably associated with the presence of MVI (), satellite nodules () and LNM (Figure S1A,S1B) in the training set. The formulas for the imaging scores in the LASSO models are listed in Table S1. Similarly, the LASSO logistic regression algorithm was also used to detect the clinical features associated with LNM in the training set (Figure S1C,S1D), and the formula for clinical scores in the LASSO model is listed in Table S1.

Table 3

Univariate and multivariate analysis of MVI presence based on clinical features in the training set

Risk factor	Univariate analysis				Multivariate analysis
	OR	95% CI	P value		OR	95% CI	P value
Age, year
<65	1.0
≥65	1.1	0.4–2.8	0.879
Gender
Male	1.0
Female	1.3	0.7–2.6	0.448
Hepatitis B virus
Negative	1.0
Positive	1.0	0.4–2.6	0.944
Tumor location
Right liver lobe	1.0
Left liver lobe	1.5	0.6–3.6	0.369
Other location	1.0	0.1–11.8	0.968
No. of tumors
Solitary	1.0
Multiple	1.1	0.5–2.3	0.804
Size, cm
<5	1.0				1.0
≥5	3.0	1.4–6.5	0.006		1.7	0.5–5.1	0.370
AFP, ng/mL
<400	1.0				1.0
≥400	3.4	1.3–8.5	0.010		2.6	0.8–8.9	0.118
CEA, ng/mL
<5	1.0
≥5	1.4	0.5–3.6	0.488
CA19-9, ng/mL
<37	1.0
≥37	0.6	0.2–1.5	0.283
ALT, U/L
<40	1.0
≥40	0.7	0.3–1.7	0.490
AST, U/L
<40	1.0
≥40	1.3	0.6–3.1	0.534
GGT, U/L
<50	1.0
≥50	1.2	0.6–2.4	0.686
Total bilirubin, μmol/L
<17.1	1.0
≥17.1	1.0	0.4–2.3	0.939
Direct bilirubin, μmol/L
<6.8	1.0
≥6.8	0.6	0.2–1.6	0.281
Albumin, g/L
<35	1.0
≥35	1.4	0.1–16.2	0.775
NLR
<2.19	1.0				1.0
≥2.19	2.2	1.0–4.7	0.049		0.9	0.2–3.6	0.901
PLR
<76.74	1.0				1.0
≥76.74	4.6	1.8–11.8	0.001		4.8	1.1–19.7	0.031
LMR
<3.03	1.0				1.0
≥3.03	0.4	0.2–1.0	0.047		0.3	0.1–1.3	0.098
AAPR
<0.53	1.0				1.0
≥0.53	2.9	1.3–6.4	0.008		4.8	1.3–17.3	0.017
APRI
<0.16	1.0
≥0.16	0.6	0.3–1.3	0.212
ANRI
<14.73	1.0
≥14.73	1.4	0.6–3.4	0.416

MVI, microvascular invasion; OR, odds ratio; CI, confidence interval; AFP, α‐fetoprotein; CEA, carcinoembryonic antigen; CA 19-9, carbohydrate antigen 19-9; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyltranspeptidase; NLR, neutrophil to lymphocyte ratio; PLR, platelet to lymphocyte ratio; LMR, lymphocyte to monocyte ratio; AAPR, albumin-to-alkaline phosphatase ratio; APRI, aspartate aminotransferase to platelet ratio index; ANRI, aspartate aminotransferase to neutrophil ratio index.

Table 4

Univariate and multivariate analysis of presence of satellite nodules based on clinical features in the training set

Risk factor	Univariate analysis				Multivariate analysis
	OR	95% CI	P value		OR	95% CI	P value
Age, year
<65	1.0
≥65	0.7	0.2–2.7	0.617
Gender
Male	1.0
Female	1.1	0.5–2.7	0.833
Hepatitis B virus
Negative	1.0
Positive	0.8	0.2–2.7	0.715
Tumor location
Right liver lobe	1.0
Left liver lobe	1.7	0.6–4.8	0.315
Other location	1.0	0.1–12.1	0.992
No. of tumors
Solitary	1.0
Multiple	1.0	0.4–2.6	0.944
Size, cm
<5	1.0				1.0
≥5	5.8	2.4–14.0	<0.001		4.3	1.1–16.1	0.003
AFP, ng/mL
<400	1.0
≥400	2.0	0.8–4.7	0.116
CEA, ng/mL
<5	1.0				1.0
≥5	2.9	1.0–8.3	0.045		12.3	1.4–107.1	0.023
CA19-9, ng/mL
<37	1.0
≥37	1.3	0.5–3.3	0.533
ALT, U/L
<40	1.0
≥40	0.6	0.2–1.5	0.253
AST, U/L
<40	1.0
≥40	1.0	0.4–2.5	0.993
GGT, U/L
<50	1.0
≥50	1.1	0.5–2.4	0.805
Total bilirubin, μmol/L
<17.1	1.0
≥17.1	0.5	0.2–1.4	0.193
Direct bilirubin, μmol/L
<6.8	1.0
≥6.8	0.6	0.2–1.7	0.320
Albumin, g/L
<35	1.0
≥35	1.8	0.2–15.2	0.991
NLR
<2.19	1.0
≥2.19	1.3	0.6–2.8	0.558
PLR
<76.74	1.0
≥76.74	3.0	1.0–9.3	0.051
LMR
<3.03	1.0
≥3.03	0.8	0.4–2.0	0.702
AAPR
<0.53	1.0
≥0.53	1.1	0.5–2.5	0.782
APRI
<0.16	1.0
≥0.16	1.0	0.5–2.3	0.966
ANRI
<14.73	1.0
≥14.73	0.9	0.3–2.3	0.747

OR, odds ratio; CI, confidence interval; AFP, α‐fetoprotein; CEA, carcinoembryonic antigen; CA 19-9, carbohydrate antigen 19-9; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyltranspeptidase; NLR, neutrophil to lymphocyte ratio; PLR, platelet to lymphocyte ratio; LMR, lymphocyte to monocyte ratio; AAPR, albumin-to-alkaline phosphatase ratio; APRI, aspartate aminotransferase to platelet ratio index; ANRI, aspartate aminotransferase to neutrophil ratio index.

Figure 4

Selection of imaging features using the LASSO logistic regression algorithm. The penalization coefficient λ in the LASSO model was tuned using 10-fold cross-validation and minimum criterion for presence of MVI (A) and satellite nodules (B). The lower X-axis shows log (λ), while the top X-axis indicates the number of predictors for the given log (λ). The Y-axis indicates binomial deviance. Red dots represent average misclassification errors for every model with given λ, and the vertical bars indicate the upper and lower values of the misclassification errors. The vertical black lines define the optimal λ, which provides its best fit to the data. As a result, an optimal λ of 0.031, with log (λ) =−3.471, was selected for MVI prediction and an optimal λ of 0.028, with log (λ) =−3.589, was selected for satellite nodules prediction. LASSO coefficient profiles of the imaging characteristics of MVI (C) and satellite nodules (D). MVI, microvascular invasion; LASSO, least absolute shrinkage, and selection operator.

Performance and validation of the nomograms

According to the results of the multivariate logistic regression and LASSO regression, clinical risk factors and imaging features for the prediction of each outcome were identified, and then integrated nomograms were constructed by combining the clinical risk factors and imaging models for MVI (), satellite nodules () and LNM (Figure S2). The C-index for the nomogram predicting MVI was 0.857 (95% CI: 0.787–0.927) in the training set and 0.795 (95% CI: 0.674–0.916) in the validation set (); the C-index for the prediction of satellite nodules was 0.919 (95% CI: 0.858–0.980) in the training set and 0.883 (95% CI: 0.740–0.906) in the validation set (); and the C-index for the nomogram predicting LNM was 0.872 (95% CI: 0.711–0.900) in the training set and 0.666 (95% CI: 0.605–0.751) in the validation set (Figure S3). We also did subgroup analysis for validating the constructed model for classical type of CHCC-CCA, the C-index for the prediction of MVI, satellite nodules and LNM was 0.788, 0.885 and 0.952 respectively (Figure S4A-S4C). Calibration plots of the training and validation sets all graphically showed good agreement between the actual risk confirmed by pathological examination and the predicted risk for MVI (), satellite nodules () and LNM (Figure S5) in both the training and validation sets. The decision curves of the combined nomograms, clinical model and imaging model are shown in for MVI, for satellite nodules and Figure S6 for LNM, indicating that our nomograms offered more benefits than the treat-all or treat-none schemes.

Figure 5

Nomogram for predicting the presence of MVI (A) and satellite nodules (B) preoperatively. Points are assigned for PLR, AAPR and imaging model for predicting the presence of MVI. Points are assigned for CEA, size, and imaging model for predicting the presence of satellite nodules. The score for each value is assigned by drawing a line upward to the points line, and the sum of score is plotted on the Total points line. MVI, microvascular invasion; PLR, platelet to lymphocyte ratio; AAPR, albumin-to-alkaline phosphatase ratio; CEA, carcino-embryonic antigen.

Figure 6

Receiver operating curve for prediction of MVI (A) and satellite nodules (B) using the combination nomogram in the training set and validation set. ROC, receiver operating characteristic; MVI, microvascular invasion.

Figure 7

The calibration curves of the combination nomogram for prediction of MVI and satellite nodules in the training set (A,C) and the validation set (B,D). Vertical axis: the observed probability of MVI and satellite nodules probabilities; horizontal axis: the nomogram predicted MVI and satellite nodules probabilities. MVI, microvascular invasion.

Figure 8

The decision curves of the combination nomogram, clinical model, and imaging model for prediction of MVI (A) and satellite nodules (B) in overall patients. Vertical axis: the net benefit; horizontal axis: the threshold probability at a range of 0.0 to 1.0. The gray line represents the decision curve of the assumption that all patients suffer from MVI or satellite nodules; the black line represents the decision curve of the assumption that no patients suffer from MVI or satellite nodules. MVI, microvascular invasion.

We observed C-indexes corresponding to the area of the ROC curves among different models for predicting MVI, satellite nodules and LNM and found that the combined nomograms were superior to those of both the clinical and imaging models for the prediction of all outcomes in the training set and validation set. According to the ROC analysis, our constructed nomograms had improved predictive value for MVI, satellite nodules and LNM and a higher threshold probability compared to the other models. The specific performances of the combined nomogram, clinical model and imaging model are summarized in .

Table 5

Performances of the clinical model, imaging model and combination nomogram

Outcome	Group	Model	C-index (95% CI)	Sensitivity	Specificity	PPV	NPV
MVI	Training set	Nomogram	0.857 (0.787–0.927)	0.854	0.765	0.719	0.881
		Clinical model	0.696 (0.607–0.784)	0.854	0.441	0.519	0.811
		Imaging model	0.802 (0.720–0.883)	0.708	0.824	0.739	0.800
	Validation set	Nomogram	0.795 (0.674–0.916)	0.815	0.714	0.733	0.800
		Clinical model	0.522 (0.373–0.672)	0.667	0.500	0.563	0.609
		Imaging model	0.839 (0.734–0.945)	0.963	0.607	0.703	0.944
Satellite nodules	Training set	Nomogram	0.919 (0.858–0.980)	0.905	0.828	0.543	0.975
		Clinical model	0.751 (0.627–0.874)	0.762	0.688	0.356	0.928
		Imaging model	0.914 (0.848–0.980)	0.810	0.903	0.654	0.955
	Validation set	Nomogram	0.883 (0.740–0.906)	1.000	0.544	0.300	1.000
		Clinical model	0.783 (0.641–0.891)	0.889	0.652	0.333	0.968
		Imaging model	0.818 (0.702–0.933)	1.000	0.609	0.333	1.000
LNM	Training set	Nomogram	0.872 (0.711–0.900)	1.000	0.842	0.288	1.000
		Clinical model	0.630 (0.628–0.678)	1.000	0.805	0.253	1.000
		Imaging model	0.683 (0.625–0.750)	1.000	0.719	0.196	1.000
	Validation set	Nomogram	0.666 (0.605–0.751)	0.666	0.881	0.235	0.980
		Clinical model	0.654 (0.632–0.664)	0.657	0.877	0.243	0.887
		Imaging model	0.617 (0.521–0.724)	0.667	0.809	0.160	0.978

CI, confidence interval; PPV, positive predictive value; NPV, negative predictive value; MVI, microvascular invasion; LNM, lymph node metastasis.

Prognostic value of the predicted MVI risk status and satellite nodule status

To evaluate the role of the constructed nomograms in the prediction of RFS and OS, we divided patients into high-risk and low-risk groups based on the median risk score calculated by the nomograms for the prediction of MVI and satellite nodules. The median risk scores for MVI and satellite nodules were −0.546 and −2.539, respectively. Survival analysis showed that patients with a high risk of MVI (P=0.003) () and satellite nodules (P=0.042) () had significantly worse RFS than their counterparts. Although there was no statistical difference observed in OS of different risk groups of MVI (P=0.057) (), patients with a high risk of having satellite nodules presented with worse OS than patients from the low-risk group (P=0.006) (). Furthermore, the performances of the predicted risk status of presenting MVI and satellite nodules for the prediction of OS at 1, 2 and 3 years were 0.755, 0.659 and 0.587, and 0.769, 0.722 and 0.681, respectively ().

Figure 9

The prognostic value of the combination nomogram of MVI and satellite nodules stratified by different risk groups. Recurrence free survival curves were drawn by the predicted risk status of MVI (A) and satellite nodules (B) by the combination nomogram. OS curves were drawn by the predicted risk status of MVI (C) and satellite nodules (D). ROC curves for predicted risk status of MVI (E) and satellite nodules (F) for predicting OS in the total cohort. MVI, microvascular invasion; ROC, receiver operating characteristic; OS, overall survival.

Discussion

CHCC-CCA has several overlapping imaging features with ICC and HCC because of hepatocellular or cholangiocellular differentiation in variable proportion, which could lead to misdiagnosis before surgery. As a result, the misdiagnosis could not guarantee the optimal treatment choice for patients with CHCC-CCA. Recently, abundant researches have proposed diagnostic algorithm of CHCC-CCA. Gigante et al. (24) developed a two-step strategy combining imaging and tumor biopsy, achieving satisfactory diagnostic performance. Moreover, according to the LI-RADS (25), tumor with imaging features of malignancy but are not specific for HCC are categorized as LR-M, and of definitive features for HCC are categorized as LR-4/5, which means that if CHCC-CCA could be categorized as LR-M, it can be correctly diagnosed with the help of biopsy before surgery. It was reported that CHCC-CCAs were categorized as LR-M and LR-5/4 in 61.4% and 37.1% respectively (8). Using LI-RADS, a substantial proportion of CHCC-CCAs can be categorized as LR-M. In this scenario, given that advances in the preoperative diagnosis of CHCC-CCA could be conveniently applied in the clinical cases, our nomograms for predicting MVI, satellite nodules and LNM preoperatively could further facilitate treatment strategy optimization for a subset of patients who can have definitive diagnosis using strategy of combined imaging and biopsy, or patients who are categorized as LR-M.

The intrinsic properties of CHCC-CCA differ from those of HCC and ICC. Moreover, its OS rate has been reported to be even lower than that of ICC (26-28). As a result, it is critical to separately map prediction models for strong prognostic parameters, including MVI, satellite nodules and LNM. Our results demonstrated that the combined nomogram models constructed for MVI, satellite nodules and LNM had satisfactory prognostication performance in both the training and validation sets and showed generally higher predictive value than the clinical and imaging models.

MVI is defined as the presence of tumor cells within a vascular lumen that is visible only by microscopy (29,30). In contrast with MVI, satellite nodules are separate micrometastatic lesions considered to be direct malignant cell invasion of the hepatic parenchyma (31). Activation of epithelial-mesenchymal transition (EMT) transcriptional programs has been proven to be an essential pathogenic step leading to the development of MVI and satellitosis (32). Since both are associated with aggressive tumor behavior, they probably have overlapping prognostic outcomes to some extent, which can be inferred from the multiple studies in which MVI was reported to be an independent risk factor for recurrence, similar to satellite nodules (33,34). Although there has been no precise classification system for MVI burden, several studies have identified the necessity of proposing novel stage criteria, as they found that the severity of MVI was highly associated with different prognoses (35,36). Therefore, adverse factors for predicting different degrees of MVI are of great importance and should be analyzed further. LNM is another important predictor of a poor prognosis in CHCC-CCA patients. Hilar LNM is rarely observed in HCC patients but is often observed in patients with CHCC-CCA and ICC (10). An accurate understanding of lymph node status enables surgeons to better tailor treatment. Although our established combined nomogram showed high C-indexes for LNM prediction in both the training and validation sets, these findings should be further validated in larger cohorts.

The proposed nomograms integrated clinical and imaging risk factors and were demonstrated to be more reliable for the prediction of MVI, satellite nodules and LNM than the nomograms constructed from clinical factors or imaging factors alone. A recent study integrated clinical risk factors and laboratory blood indicators to investigate the diagnostic accuracy of MVI and satellite nodules for CHCC-CCA, yielding C-indexes of 0.826 and 0.771, respectively (37), which were lower than those of the nomograms we built. In another study, imaging features of intratumoral fat deposition and irregular peritumoral enhancement combined with the AFP level for the prediction of MVI yielded a specificity of 98.2% but a sensitivity of 12.5% (38).

The inflammatory response plays a decisive role in cancer initiation, malignant transformation and MVI by promoting epigenetic changes, cell proliferation, angiogenesis, and cell invasion (39). In our study, the level of PLR was significant serum inflammatory markers of MVI in CHCC-CCA, which was consistent with prior studies. A meta-analysis showed that elevated PLR suggested poor OS in HCC, with a pooled HR of 1.63 (40). Albumin is also a useful marker for determining the inflammatory response, and alkaline phosphatase hydrolase enzymes are primarily found in the liver, bile duct and bone. In our nomogram, AAPR was an independent predictor of MVI, concordant with the findings reported in the study by Zhang et al. (41), who identified that AAPR was a strong indicator of OS in CHCC-CCA, with an optimal threshold value of 0.43. Similarly, utilizing these inflammatory indexes in our nomogram could improve its predictive power. Interestingly, PLR and AAPR were not independent risk factors for the prediction of satellite nodules in our study, probably because of the smaller number of events in satellite nodules compared with that in MVI, which led to weakened statistical power.

Imaging can help to visualize the entire profile of the relative proportions of tumor features. Several studies have reported that lesions harboring a hypervascular pattern had a larger proportion of HCC features and less fibrotic stroma (42,43), while lesions with a targeted appearance, as defined by LI-RADS, were more likely to have a larger proportion of CC features and fibrotic stroma (38). The intrinsic heterogeneity of CHCC-CCA histology contributes to the discordance in survival outcomes. This idea was based on the evidence reported by Park et al. (42), demonstrating that CHCC-CCA patients with hypervascular features had better survival outcomes than CHCC-CCA patients with nonhypervascular features; simultaneously, patients categorized as having LR-M or LR-4/5 were found to have significant differences in survival outcomes (7,8). The multiple imaging biomarkers selected by LASSO logistic regression in our study yielded satisfactory prediction performance. However, interobserver variability hampers the quality of imaging feature evaluations, and radiomics has begun to emerge as a useful tool to predict the status of MVI and LNM in HCC (44,45); however, such a method has not yet been applied to CHCC-CCA. In the future, the establishment of the genomic, transcriptomic, proteomic and radiomic landscapes of CHCC-CCA is desirable for a better understanding of the progression of CHCC-CCA and adverse prognostic factors.

Unfortunately, there is no globally accepted standard first-line therapy for CHCC-CCA, and comparative data on different therapeutic tools are sparse. Thus, the nomograms we constructed for predicting the most predominant prognostic factors might serve as an effective tool to guide appropriate clinical treatment selection. Evidence from a multicenter analysis showed that compared with resection, the outcomes of patients with CHCC-CCA who underwent transplantation were superior (46); however, they exhibited inferior 5-year survival outcomes compared to those with HCC who underwent transplantation (47). Moreover, patients with a high MVI status were reported to be unsuitable for transplantation since unsatisfactory OS outcomes were observed compared with patients without MVI (48). Therefore, preoperative estimated models could facilitate clinicians in recruiting candidates who could obtain the most benefit from transplantation. For patients who present with MVI or satellitosis, a larger resection margin could potentially clear intrahepatic metastases. In this scenario, RFA, as an alternative ablative therapy of hepatic resection, showed less favorable surgical outcomes due to inferior radical tumor clearance. Neoadjuvant therapy for ICC with LNM is expected to improve survival (49), providing supportive evidence that patients could receive neoadjuvant therapy for CHCC-CCA. There is also accumulating evidence suggesting that aggressive surgical treatment, including lymph node dissection, can prolong the survival of patients with CHCC-CCA (28). In this scenario, successful preoperative prediction of LNM risk could result in improved postoperative survival outcomes.

Our study had some limitations that should be mentioned. First, since our validation set was derived from a single institution, our available sample size was relatively small. Hence, analysis based on a larger cohort and external validation in other centers is necessary. Second, a prospective study is required to further confirm the reliability of our nomograms. Third, we did not categorize CHCC-CCA into different subtypes according to the latest WHO classification system; more specific prognostication among various types should be conducted to further elucidate the biological behavior of CHCC-CCA. Finally, since qualitative imaging features suffer from some limitations, including interobserver variability and lack of external validation, studies incorporating radiomics or genomics into predictive nomograms should be conducted in the future.

Conclusions

Nomograms incorporating clinical factors and imaging features achieved satisfactory performance in individualized prediction of preoperative status of MVI, the presence of satellite nodules and LNM, allowing treatment strategy optimization in patients with CHCC-CCA. Furthermore, these investigations need to be validated in more prospective studies.

The article’s supplementary files as