Combining preoperative clinical and ultrasound radiomic features to predict the predominant component of combined hepatocellular carcinoma-cholangiocarcinoma

Abstract

Background

As a rare subtype of primary liver cancer that contains both hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA) components, combined hepatocellular carcinoma-cholangiocarcinoma (cHCC-CCA) is characterized by strong invasiveness and poor prognosis. The proportion of HCC and CCA components is closely related to the prognosis of patients. This study aimed to develop a combined model integrating preoperative ultrasound radiomic features and clinical data to non-invasively predict the predominant component in cHCC-CCA.

Methods

The study included 60 histopathologically confirmed cHCC-CCA patients, divided into high hepatocellular carcinoma percentage (HCC%) group (n = 33, HCC% > 50) and low HCC% group (n = 27, HCC% ≤ 50). A predominant component prediction model for cHCC-CCA was developed using radiomic features extracted from four distinct region of interest (ROI) to identify the optimal radiomic model: tumor alone (ROI_tumor), tumor plus 5 mm peritumoral margin (defined as ROI_tumor+peritumor 1), tumor plus 10 mm peritumoral margin (defined as ROI_tumor+peritumor 2), and tumor plus 20 mm peritumoral margin (defined as ROI_tumor+peritumor 3). Univariate and multivariate logistic regression analyses were performed to evaluate the predictive value of preoperative clinical parameters as potential indicators of the predominant component predictors. Ultimately, a combined model integrating the optimal radiomic features and clinical data was constructed, with the final predictive model visualized using nomogram.

Results

The model developed based on ROI_tumor+peritumor 2 demonstrated optimal predictive performance and was defined as the optimal radiomic model [training set area under the curve (AUC) = 0.853, test set AUC = 0.806, overall set AUC = 0.832]. Through multivariate logistic regression analysis, we identified two preoperative indicators: tumor maximum diameter and total bilirubin. Compared with both the optimal radiomic model and the clinical model, the combined model demonstrated superior predictive capability for the predominant component in preoperative cHCC-CCA patients (training set AUC = 0.877, 95% CI: 0.775–0.980; test set AUC = 0.847, 95% CI: 0.657–1.00; overall set AUC = 0.851, 95% CI: 0.756–0.945). Decision curve analysis demonstrated that the combined model exhibited favorable clinical utility.

Conclusions

Combination of ultrasound-based radiomics with tumor maximum diameter and total bilirubin levels holds significant value for the preoperative identification of HCC% in cHCC-CCA.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-026-04966-4.

Introduction

Primary liver cancer is histologically categorized into hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), and combined hepatocellular carcinoma-cholangiocarcinoma (cHCC-CCA). Among these, cHCC-CCA represents a rare and highly aggressive hepatic malignancy, with reported incidence rates ranging from 0.4% to 14.2% [1–3]. It is primarily characterized by the coexistence of hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA) pathological differentiation within the same tumor [4]. Notably, the incidence and mortality rates of cHCC-CCA have shown an increasing trend in recent years [2, 5]. Currently, the standard therapeutic approach for cHCC-CCA involves surgical resection accompanied by lymph node dissection [1, 6]. cHCC-CCA has a composite histological nature. It therefore shares similar imaging characteristics with HCC and ICC, which frequently leads to misdiagnosis as either of these two subtypes [6, 7]. Consequently, the unique histopathological features, biological behavior, radiological presentation, and diagnostic challenges of cHCC-CCA have also attracted considerable attention [8].

Recent investigations have increasingly focused on identifying prognostic factors for cHCC-CCA, with the relative proportions of HCC and CCA components recognized as critical determinants of patient outcomes. Evidence indicates that a higher CCA component is associated with poorer prognosis in cHCC-CCA patients [7, 9], whereas a predominance of the HCC component correlates with improved postoperative overall survival [10]. Despite advancements in medical technology, studies indicate that the five-year post-operative recurrence rate for cHCC-CCA remains as high as 80% or more [11]. Controversy persists regarding the prognosis and management of cHCC-CCA; appropriately identifying prognostic factors would facilitate risk stratification and accelerate individualized management of cHCC-CCA [12]. These findings highlight the importance of preoperative assessment of the HCC and CCA component ratios, which could enhance prognostic accuracy and improve treatment outcomes. Currently, the predominant component of cHCC-CCA is mainly determined through invasive procedures, such as biopsy or surgical resection. Given the significant trauma and risks associated with these methods, the development of non-invasive techniques for the preoperative prediction of the predominant component is critically important.

Imaging examinations are routinely employed as part of the pre-treatment evaluation for patients with cHCC-CCA, allowing comprehensive visualization of the tumor. Radiomics utilizes high-throughput methodologies to extract sophisticated quantitative features from imaging data, thereby capturing critical phenotypic information [13]. Current studies suggest that radiomics analysis is valuable for predicting microvascular invasion (MVI), the presence of satellite nodules, and lymph node metastasis (LNM) in cHCC-CCA [14–16]. These findings imply that radiomics may represent a viable approach for the preoperative evaluation of the relative proportions of HCC and CCA components. Nevertheless, the literature addressing the preoperative prediction of the predominant tumor component in cHCC-CCA patients remains limited. One investigation developed a predictive model for the percentage of HCC (HCC%) in cHCC-CCA patients by integrating MRI-derived habitat volume fractions, conventional imaging features, clinical parameters, and RNA sequencing data [12]. Another study revealed that MRI-based Liver Imaging Reporting and Data System (LI-RADS) classification and elevated serum carcinoembryonic antigen (CEA) levels (> 5 ng/mL) serve as significant predictors of HCC% in cHCC-CCA patients [8]. Collectively, these studies underscore the considerable utility of imaging characteristics in the preoperative assessment of HCC% in patients with cHCC-CCA. To date, no research has explored the application of ultrasound imaging for predicting the predominant component in cHCC-CCA. Consequently, developing a convenient, accurate, and non-invasive preoperative method based on ultrasound imaging to predict the predominant component in cHCC-CCA patients could substantially enhance clinical decision-making.

Previous research has demonstrated that analyzing imaging features in peritumoral tissues can serve as a tool for predicting MVI and the presence of satellite nodules in cHCC-CCA patients. For instance, Zhang et al. utilized contrast-enhanced CT imaging features from 60 cHCC-CCA patients to predict MVI, finding that peritumoral characteristics were crucial for this prediction [14]. Similarly, Huang et al. analyzed radiological imaging features (CT or MRI) from 69 cHCC-CCA patients and demonstrated that peritumoral regions also provide critical information regarding tumor MVI [3]. Furthermore, Wang et al. integrated clinical risk factors with MRI imaging features to develop a nomogram aimed at the preoperative prediction of MVI, satellite nodules, and LNM [16]. Accordingly, this study constructed ultrasound radiomic models using the tumor and incremental peritumoral margins, aiming to identify the optimal region for predicting the predominant component. To explore the impact of imaging features from both tumor and peritumoral regions on principal component prediction, this study classified cHCC-CCA patients into high and low HCC% groups on the basis of histopathological examination. The study’s objective was to construct ultrasound radiomic models utilizing various region of interest (ROI) and evaluate the predictive utility of preoperative clinical parameters alongside ultrasound radiomic features in the predominant component of cHCC-CCA.

Materials and methods

Patient selection

The Ethics Committee of the First Affiliated Hospital of Guangxi Medical University approved this study. We retrospectively collected clinical, pathological, and imaging data from patients with pathologically confirmed cHCC-CCA at the First Affiliated Hospital of Guangxi Medical University between January 2016 and February 2025. The screening process is illustrated in Fig. 1. On the basis of the pathological results, the patients were categorized into high and low HCC% groups, with 50% used as the threshold. The specific inclusion and exclusion criteria were as follows:

Fig. 1.

Flow diagram of the study population. Abbreviations: cHCC-CCA: combined hepatocellular carcinoma-cholangiocarcinoma, HCC%: percentage of hepatocellular carcinoma

Inclusion criteria: (1) pathologically confirmed cHCC-CCA and (2) complete ultrasound imaging records. Exclusion criteria: (1) incomplete clinical data. (2) poor ultrasound image quality. (3) pathology reports lacking HCC and CCA component proportions, and (4) prior treatment (surgery, ablation, targeted immunotherapy, interventional procedures, or chemotherapy) before ultrasound examination.

Ultrasound image acquisition and region of interest delineation

Ultrasound examinations and image storage of hepatic lesions were conducted by two sonographers: one with a minimum of five years and the other with at least ten years of clinical experience in abdominal ultrasound diagnostics. In cases of disagreement, the final decision was reached through discussion between the two physicians. Neither the physician had access to the basic patient information nor the pathological findings associated with each ultrasound image. Multiple diagnostic ultrasound systems were utilized, mainly including the Resona 7 and Resona R9 Pro (Mindray, China), Logiq E9 (GE Healthcare, USA), and Philips EPIQ 5 (Philips Medical Systems, USA). Abdominal ultrasound probes operate at frequencies ranging from 2 to 5 MHz. Ultrasound images showing the largest cross-sectional view of each lesion were selected for acquisition.

3D Slicer software (version 5.6.1) was used to delineate the liver tumor and the ROI in the liver parenchyma surrounding the tumor at distances of 5 mm (defined as peritumor 1), 10 mm (defined as peritumor 2), and 20 mm (defined as peritumor 3). Following confirmation of the ROI_tumor, the peritumoral contour was automatically expanded from this baseline, ensuring the exclusion of vessels, gas, bone, and bile ducts. If the expanded ROI extends beyond the liver parenchyma, manual adjustments are required to exclude areas outside the liver tissue. For patients with multiple liver lesions, we selected the lesion with the maximum diameter for evaluation.

Radiomic features extraction and model construction

The Python package “PyRadiomics” was employed to extract 1,409 radiomic features per ROI, including shape, histogram, and texture-based characteristics. These include morphological, first-order, second-order (texture features), filter transform, and wavelet transform features. First, to evaluate the stability and reproducibility of radiomic features, we randomly selected 30 lesion images for another sonographer to redraw the ROI and re-extract the features. Interobserver variability was analyzed via the intraclass correlation coefficient (ICC), which retains features with high stability (ICC > 0.75). Subsequently, hypothesis testing was performed. The Wilcoxon signed-rank test was employed to select features associated with the cHCC-CCA predominant component, followed by further feature selection via least absolute shrinkage and selection operator (LASSO) regression with ten-fold cross-validation. Finally, ultrasound radiomics models were constructed via logistic regression on the basis of features extracted from different ROI regions.

SHAP model interpretability analysis

To ensure the reliability of the radiomic model, we employed shapley additive explanations (SHAP) to enhance model interpretability, aiming to quantify the contribution of individual variables to the principal component prediction results for cHCC-CCA. SHAP is a model interpretation method based on the Shapley value in game theory, and the size of the SHAP value indicates the importance of a feature in the model’s prediction [17]. Additionally, the SHAP value reveals each feature’s specific contribution to individual predictions, thereby enhancing the model’s transparency and interpretability.

Clinical data processing and combined model construction

The correlations between clinical variables and the predominant components of cHCC-CCA were assessed via univariate logistic regression analysis. Variables meeting P < 0.2 were identified as potential predictors and included in multivariate logistic regression analysis. The clinical variables considered included patient gender, age, tumor size, routine blood tests, liver function, T-cell count, alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA199), protein induced by vitamin K absence or antagonist-Il (PIVKA-II), and hepatitis B infection status. Finally, we constructed a combined model integrating clinical risk factors with the optimal features and visualized it as a nomogram.

Model performance evaluation

This study utilized widely recognized evaluation metrics to assess the performance of the models, including the area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SENS), specificity (SPEC), positive predictive value (PPV), negative predictive value (NPV), Matthew’s correlation coefficient (MCC), and F1 score. Additionally, the efficacy of the combined model was further examined through the construction of confusion matrices, as well as through calibration curve analysis and decision curve analysis (DCA). Furthermore, clinical impact curve analysis demonstrated the clinical efficacy of the combined model.

Statistical analysis

T tests or Mann‒Whitney tests were employed for quantitative data, with the results expressed as the means ± standard deviations (SDs). Chi-square tests or Fisher’s exact tests were used for qualitative data, presented as n (%). Diagnostic performance was assessed via ROC curve analysis, with AUC values compared via the DeLong test. All the statistical analyses were carried out via R software (version 4.3.3). P < 0.05 were considered statistically significant.

Results

Patient characteristics

A total of 60 patients were enrolled in this study, with the majority presenting solitary lesion. All patients were stratified into the high HCC% group (HCC% > 50, n = 33) and the low HCC% group (HCC% ≤ 50, n = 27). In the high HCC% group, 28 patients (84.85%) had solitary lesion, while 27 patients (77.78%) in the low HCC% group were found to have solitary lesion. Comparative analysis of clinical parameters between the two groups revealed that only the tumor maximum diameter differed significantly (Table 1).

Table 1.

Clinical and pathological characteristics of patients with cHCC-CCA

Clinical parameter	HCC%		P value
	High (n = 33)	Low (n = 27)
Age (y)	51.88 ± 10.99	52.52 ± 9.71	0.812
Gender
male, n(%)	28 (84.85)	21 (77.78)	0.520
HBeAg-Positive, n(%)	24 (80.00)	17 (77.27)	1.000
Tumor maximum diameter (mm)	45.91 ± 32.22	67.59 ± 36.68	0.012
Number of lesions
Solitary, n (%)	28 (84.85)	21 (77.78)	0.712
Liver functional parametcrs
Total bilirubin (umol/L)	18.29 ± 18.10	12.58 ± 8.48	0.054
Total Protein (g/L)	69.62 ± 7.36	70.60 ± 8.64	0.662
ALB (g/L)	37.87 ± 5.36	39.98 ± 9.27	0.573
ALT (U/L)	108.47 ± 257.47	75.87 ± 172.06	0.213
AST (U/L)	106.72 ± 188.95	59.17 ± 58.56	0.389
GGT (U/L)	82.49 ± 107.56	91.34 ± 82.92	0.484
ALP (U/L)	99.09 ± 40.94	105.74 ± 49.42	0.663
Laboratory indicators of inflammation
NEU%	4.79 ± 3.14	0.60 ± 0.16	0.533
LYM%	1.69 ± 0.82	1.60 ± 0.57	0.739
MONO%	0.55 ± 0.27	0.66 ± 0.32	0.140
NLR	4.71 ± 7.22	3.83 ± 4.14	0.536
PLR	148.87 ± 112.44	140.95 ± 75.91	0.629
LMR	3.51 ± 1.94	2.73 ± 1.16	0.104
AFP, n(%)			1.000
≤ 400 ng/ml	20 (62.50)	15 (65.22)
﹥400 ng/ml	12 (37.50)	8 (34.78)
PIVKA-II, n(%)			0.780
≤ 40 mAU/ml	12 (40.00)	10 (45.45)
﹥40 mAU/ml	18 (60.00)	12 (54.55)
CA199, n(%)			1.000
≤ 37 U/ml	22 (68.75)	16 (72.73)
﹥37 U/ml	10 (31.25)	6 (27.27)
CEA, n(%)			0.639
≤ 5 ng/ml	3 (9.78)	1(4.76)
﹥5 ng/ml	28 (90.32)	20 (95.24)

Abbreviations: cHCC-CCA combined hepatocellular carcinoma-cholangiocarcinoma, ALB Albumin, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT Gamma-Glutamyl transferase, ALP Alkaline phosphatase, NEU% Neutrophil granulocytes ratio, LYM% Lymphocyte ratio, MONO% Monocyte ratio, NLR Neutrophil/lymphocyte ratio, PLR Platelet/lymphocyte ratio, LMR Lymphocyte/monocyte ratio, AFP Alpha-fetoprotein, PIVKA-II Protein induced by vitamin K absence or antagonist-Il, CA199 Carbohydrate antigen 199, CEA Carcinoembryonic antigen

P < 0.05 was considered statistically significant

Univariate and multivariate logistic regression analyses of clinical characteristics

Univariate and multivariate logistic regression analyses were conducted on all the baseline data of the high and low HCC% groups. Multivariate analysis identified tumor maximum diameter (P = 0.036) and total bilirubin (P = 0.034) as independent significant risk factors for high HCC composition (Table 2). These factors are subsequently incorporated into the combined model construction.

Table 2.

Uni-/multivariate logistic regression analysis of clinical data

Characteristics	Univariate analysis		Multivariate analysis
	P	OR (95%CI)	P	OR (95%CI)
Age (y)	0.810	0.99 (0.95 ~ 1.04)
Tumor maximum diameter (mm)	0.026	0.98 (0.97 ~ 0.99)	0.036	0.92 (0.86 ~ 0.99)
NEU%	0.781	0.98 (0.82 ~ 1.17)
LYM%	0.630	1.21 (0.56 ~ 2.58)
MONO%	0.168	0.25 (0.04 ~ 1.79)
NLR	0.600	1.03 (0.93 ~ 1.13)
PLR	0.766	1.00 (1.00 ~ 1.01)
LMR	0.099	1.39 (0.94 ~ 2.07)	0.216	1.97 (0.67 ~ 5.77)
Total bilirubin (µmol/L)	0.182	1.05 (0.98 ~ 1.13)	0.034	1.12 (1.01 ~ 1.24)
ALB (g/L)	0.316	0.96 (0.88 ~ 1.04)
ALT (U/L)	0.601	1.00 (1.00 ~ 1.00)
AST (U/L)	0.290	1.00 (1.00 ~ 1.01)
GGT (U/L)	0.738	1.00 (0.99 ~ 1.00)
ALP (U/L)	0.582	1.00 (0.98 ~ 1.01)
T cell	0.545	0.97 (0.87 ~ 1.07)
CD4+ T cell	0.074	0.90 (0.81 ~ 1.01)	0.191	0.88 (0.74 ~ 1.06)
CD8+ T cell	0.950	1.00 (0.92 ~ 1.09)
AFP
﹥ 400 ng/ml		1.00 (Reference)
๥>400 ng/ml	0.836	1.13 (0.37 ~ 3.44)
PIVKA-II
≤ 40 mAU/ml		1.00 (Reference)
﹥40 mAU/ml	0.694	1.25 (0.41 ~ 3.80)
CA199
≤ 37 U/ml		1.00 (Reference)
﹥37 U/ml	0.753	1.21 (0.37 ~ 4.02)
CEA
≤ 5 ng/ml		1.00 (Reference)
﹥5 ng/ml	0.522	0.47 (0.05 ~ 4.82)
Hepatitis B virus
No		1.00 (Reference)
Yes	0.812	1.18 (0.31 ~ 4.49)

Abbreviations: cHCC-CCA combined hepatocellular carcinoma-cholangiocarcinoma, NEU% Neutrophil granulocytes ratio, LYM% Lymphocyte ratio, MONO% Monocyte ratio, NLR Neutrophil/lymphocyte ratio, PLR Platelet/lymphocyte ratio, LMR Lymphocyte/monocyte ratio, ALB Albumin, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT Gamma-Glutamyl transferase, ALP Alkaline phosphatase, AFP Alpha-fetoprotein, PIVKA-II Protein induced by vitamin K absence or antagonist-Il, CA199 Carbohydrate antigen 199, CEA Carcinoembryonic antigen

P < 0.05 was considered statistically significant

Selection and modeling of radiomic features

This study excluded radiomic features with ICC ≤ 0.75. Subsequently, hypothesis tests were conducted on the remaining radiomic features to identify those exhibiting statistically significant differences between the high and low HCC% groups. According to random stratified sampling principles, patients were stratified into a training set (n = 43) and a test set (n = 17) at a 7:3 ratio. The overall set comprises the total number of samples in both the training and test sets. Finally, we applied the LASSO algorithm to the training set to identify and select the most predictive features for model construction. Using consistent training and test sets, we performed the aforementioned processing for the following regions: ROI_tumor, ROI_tumor+peritumor 1, ROI_tumor+peritumor 2, and ROI_tumor+peritumor 3. The training, test, and overall set AUC values for ROI_tumor were 0.829, 0.750, and 0.814, respectively. For ROI_tumor+peritumor 1, the corresponding AUC values were 0.846, 0.750, and 0.817. For ROI_tumor+peritumor 3, the values were 0.774, 0.778, and 0.781. In contrast, the ROI_tumor+peritumor 2 demonstrated higher AUC values of 0.853, 0.806, and 0.832 in the training, test, and overall sets, respectively (Table 3). Additionally, we employed the DeLong test to compare the AUC values of the four radiomics models in predicting the predominant component of cHCC-CCA, and no statistically significant differences were observed (Supplementary Table 1). Ultimately, by integrating metrics, including the ACC, SENS, SPEC, PPV, NPV, MCC, and F1 score, we found that the ROI_tumor+peritumor 2 outperformed the other regions in predicting the predominant component of cHCC-CCA, establishing it as the optimal radiomic model. Within the ROI_tumor+peritumor 2, we extracted 1,409 radiomic features. After consistency testing and Wilcoxon rank-sum tests, 204 features showed statistically significant differences between the high and low HCC% groups. Subsequently, we selected the five most relevant features using the LASSO method and incorporated them into logistic regression analysis to construct the radiomic model (Fig. 2).

Table 3.

Prediction performance of different radiomics models for the predominant component in cHCC-CCA

Group	Models	AUC	ACC	SENS	SPEC	PPV	NPV	MCC	F1 score
Training set	ROItumor	0.829	0.837	0.750	0.947	0.947	0.750	0.697	0.837
	ROItumor+peritumor 1	0.846	0.791	0.625	1.000	1.000	0.679	0.651	0.769
	ROItumor+peritumor 2	0.853	0.791	0.667	0.947	0.941	0.692	0.624	0.780
	ROItumor+peritumor 3	0.774	0.744	0.792	0.684	0.760	0.722	0.479	0.776
Test set	ROItumor	0.750	0.824	1.000	0.625	0.750	1.000	0.685	0.857
	ROItumor+peritumor 1	0.750	0.765	0.889	0.625	0.727	0.833	0.537	0.800
	ROItumor+peritumor 2	0.806	0.765	0.667	0.875	0.857	0.700	0.549	0.750
	ROItumor+peritumor 3	0.778	0.765	0.778	0.750	0.778	0.750	0.528	0.778
Overall set	ROItumor	0.814	0.800	0.727	0.889	0.889	0.727	0.616	0.800
	ROItumor+peritumor 1	0.817	0.767	0.636	0.926	0.913	0.676	0.575	0.750
	ROItumor+peritumor 2	0.832	0.767	0.667	0.889	0.880	0.686	0.561	0.759
	ROItumor+peritumor 3	0.781	0.750	0.788	0.704	0.765	0.731	0.494	0.776

Abbreviations: cHCC-CCA Combined hepatocellular carcinoma-cholangiocarcinoma, ROI Region of interest, AUC Area under curve, ACC Accuracy, SENS Sensitivity, SPEC Specificity, PPV Positive predictive value, NPV Negative predictive value, MCC Matthew’s correlation coefficient

Fig. 2.

ROI_tumor+peritumor 2 radiomic features selection using the LASSO binary logistic regression model. A The LASSO regression model uses 10-fold cross-validation to select the optimal tuning parameter (λ). B Five non-zero coefficient radiomics features were obtained by λ selection. Abbreviations: ROI: region of interest, LASSO: least absolute shrinkage and selection operator

Importance of radiomic features interpreted by SHAP values

SHAP visualization offers detailed insights into how radiomic model generate predictions. The five radiomic features incorporated into the ROI_tumor+peritumor 2 model are as follows: exponential_glszm_SizeZoneNonUniformityNormalized, wavelet. HLL_glszm_LargeAreaHighGrayLevelEmphasis, wavelet.HLL_glszm_ZonePercentage, wavelet. HLH_glszm_ZonePercentage, and wavelet.HHL_gldm_SmallDependenceEmphasis. The SHAP bar plot displays the importance ranking of each feature in the radiomics model for predicting the principal component (Fig. 3A). The SHAP values indicate only the wavelet.HLL_glszm_LargeAreaHighGrayLevelEmphasis was significantly negatively correlated with high HCC% risk (Fig. 3B). We employed the shapviz package to generate waterfall and force plots, elucidating the contribution of each radiomic feature to the model’s predictive process. The contribution ranking and cumulative prediction level of the five radiomic features in predicting high HCC% risk for one cHCC-CCA patient (pathological results indicating 70% HCC and 30% CCA component ratio) yielded a prediction value of 0.839 (Fig. 3C). Yellow indicates radiomic features that increase the prediction of high HCC% risk for this patient, whereas purple indicates features that decrease the prediction value (Fig. 3D). Conversely, when the same radiomic features were used to predict high HCC% risk for another cHCC-CCA patient (pathological results indicating 30% HCC and 70% CCA component ratio), the final prediction value was only 0.111 (Fig. 3E–F). These results demonstrate that the optimized radiomic model provides strong predictive performance for the predominant components of cHCC-CCA.

Fig. 3.

Interpretation of the optimal radiomic model using the SHAP method at both global and local levels. A SHAP bar plot: ranking of the importance of 5 radiomics features based on the generated SHAP values. B SHAP bee swarm plot: correlation between the value of each radiomics feature and its SHAP value, with color indicating the feature value (yellow = high, purple = low). C-D Waterfall and force plots of a patient with high HCC% cHCC-CCA. E-F Waterfall and force plots of a patient with with HCC% cHCC-CCA. Abbreviations: cHCC-CCA: combined hepatocellular carcinoma-cholangiocarcinoma, SHAP: shapley additive explanations, HCC%: percentage of hepatocellular carcinoma

Comparison and selection of models

Following multivariate logistic regression analysis, tumor maximum diameter and total bilirubin were selected from the clinical baseline data to establish a predominant component preoperative predictive model. Subsequently, we constructed a combined model by integrating the optimal radiomic features with clinical data. The combined model achieved AUC values of 0.877 (95% CI: 0.775–0.980), 0.847 (95% CI: 0.657–1.00), and 0.851 (95% CI: 0.756–0.945) in the training, test, and overall sets, respectively (Fig. 4A–C). The DeLong test revealed statistically significant differences in AUC values between the combined model and the optimal radiomic model, as well as between the combined and the clinical models within the training set. However, in the overall set, only the difference in AUC values between the combined and clinical models were statistically significant (P < 0.05) (Table 4). Combining the ACC, SENS, SPEC, PPV, NPV, MCC, and F1 score, our results indicate that the combined model demonstrates superior predictive performance for the predominant component in preoperative cHCC-CCA patients (Supplementary Table 2).

Fig. 4.

Assessment of the performance of three models in predicting the predominant component of cHCC-CCA. A AUC for the training set (n = 43), B test set (n = 17), and (C) overall set (n = 60). The red, blue, and yellow lines represent the optimal radiomic, clinical, and combined models, respectively. AUC values are presented with 95% confidence intervals in parentheses. Abbreviations: cHCC-CCA: combined hepatocellular carcinoma-cholangiocarcinoma, AUC: area under curve

Table 4.

Comparison of ROC curves for predictive models using the Delong test

Model	Training set		Test set		Overall set
	AUC	P value	AUC	P value	AUC	P value
Radiomic	0.853		0.806		0.832
Clinical	0.675		0.819		0.725
Combined	0.877		0.847		0.851
Radiomic VS Clinical		0.396		0.619		0.437
Combined VS Radiomic		0.018		0.784		0.061
Combined VS Clinical		0.007		0.728		0.024

Abbreviations: AUC Area under curve, ROC Receiver operating characteristic

P < 0.05 was considered statistically significant

To provide personalized probability estimates and illustrate the weighting of individual features, we presented the variables included in the combined model as nomogram. The red-highlighted section details the prediction probability calculation for a cHCC-CCA patient with HCC as the predominant component (70% HCC, 30% CCA) using this nomogram score (Fig. 5A). DCA analysis indicated that the combined model delivers substantial net benefit within a reasonable threshold probability range, demonstrating its practical clinical utility (Fig. 5B). Additionally, the confusion matrix showed that the combined model achieved accuracy rates of 0.837, 0.824, and 0.800 for predicting high HCC% in cHCC-CCA patients in the training, test, and overall sets, respectively (Fig. 5C–E). Furthermore, calibration curves demonstrated statistically significant outcomes in both the training and overall sets (P < 0.05), indicating that the combined model possesses favorable calibration capabilities (Supplementary Fig. 1). When the threshold probability exceeded 70% of the predicted score probability value, the predictive model accurately identified high-risk individuals for high HCC%, with high concordance with the actual occurrence of high HCC%, confirming the model’s high clinical efficacy rate (Fig. 6).

Fig. 5.

Clinical application of the combined model. A Nomogram of the combined model predicting the predominant components of cHCC-CCA. B Net benefit of the combined model based on decision curve analysis for predicting the predominant component in cHCC-CCA. The red, blue, and green lines represent the training, test, and overall sets, respectively. C–E Confusion matrices of the combined model in the training, test, and overall sets. Abbreviations: cHCC-CCA: combined hepatocellular carcinoma-cholangiocarcinoma

Fig. 6.

Clinical impact curve of the combined model. A Training set; B Test set; C Overall set. Of the 1000 patients, the solid red line shows the total number of individuals identified as high risk at each risk threshold, while the light blue dotted line represents the number of true positives within that group

Discussion

The proportion of hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA) components in combined hepatocellular carcinoma-cholangiocarcinoma (cHCC-CCA) is closely associated with patient prognosis. Specifically, patients whose pathological profiles are dominated by the HCC component tend to experience more favorable clinical outcomes [8, 18]. Additional research has demonstrated that patients with a relatively low proportion of CCA component exhibit improved overall survival rate [2]. These findings suggest that the component ratio in cHCC-CCA can serve as a prognostic indicator. In the present study, we developed an integrative predictive model incorporating ultrasound radiomic features alongside clinical parameters to preoperatively determine the predominant tumor component in cHCC-CCA. The model exhibited moderate predictive performance, highlighting its potential utility as a non-invasive tool for the preoperative assessment of the dominant histological component in cHCC-CCA.

Microvascular invasion (MVI) and predominant components represent crucial biological characteristics of cHCC-CCA. Numerous studies have established models to predict MVI via preoperative conventional ultrasound or contrast-enhanced ultrasound examinations [19, 20]. However, research focused on predicting the predominant pathological component in cHCC-CCA, especially via ultrasound radiomics, remains limited. Previous work has integrated MRI habitat mapping with RNA sequencing to build invasive models that predict the proportions of HCC components in cHCC-CCA patients [12]. Nevertheless, this MRI-based approach is often limited by its lengthy processing times, high costs, and invasive nature. In contrast, ultrasound is widely accessible, cost-effective, convenient, and non-invasive. In this study, we developed a combined model incorporating ultrasound radiomics and clinical data to non-invasively predict the primary pathological components of cHCC-CCA preoperatively. The model demonstrated robust predictive performance (training set AUC = 0.877, test set AUC = 0.847, overall set AUC = 0.851). Furthermore, the constructed nomogram facilitates the application of this predictive model in routine clinical practice and may assist clinicians in preoperative decision-making.

Multiple studies have identified tumor maximum diameter and total bilirubin level as independent risk factors associated with MVI in HCC [21, 22] and ICC [23, 24] patients. Consistent with these findings, our results demonstrate that tumor maximum diameter and total bilirubin level serve as independent predictors of the major pathological component in cHCC-CCA. Additionally, elevated serum CEA levels have been recognized as a risk factor for HCC predominance in cHCC-CCA patients [8]. However, no significant difference in serum CEA levels was observed in the present study between groups with high and low HCC%. This discrepancy may be attributable to variations in the cut-off values used to define HCC% across studies. Moreover, prior research has suggested that the clinical and pathological features of hepatitis B virus (HBV)-associated cHCC-CCA more closely resemble those of HCC [25]. Clinical characteristics of cHCC-CCA often parallel those of HCC, including a male predominance and a high prevalence of HBV infection among patients [15, 26, 27]. Our findings align with this trend, as the male-to-female ratio was 49:11, and 78.8% of patients were infected with HBV. Nevertheless, no statistically significant differences in gender distribution or etiology were observed between the different groups, which may be attributable to our study’s relatively limited sample size.

Accumulating evidence indicates that subtle modifications in the normal tissue architecture surrounding tumors may yield significant insights into tumor biology. Similarly, Hu et al. reported that a region of interest (ROI) model encompassing the tumor and a 20 mm peritumoral margin achieved superior diagnostic accuracy for predicting MVI in HCC relative to models based solely on the tumor or the tumor plus a 10 mm peritumoral region [28]. Additionally, Feng et al. reported that a radiomics model derived from hepatobiliary phase imaging, incorporating the tumor and a 10 mm peritumoral area, outperformed a tumor-only model in predicting MVI in HCC [29]. Collectively, these findings underscore the biological importance of the peritumoral region in tumor characterization. Nonetheless, a consensus on the standardized definition of the peritumoral margin remains lacking. Our study similarly revealed that, compared with ROI_tumor, ROI_tumor+peritumor 1, and ROI_tumor+peritumor 3, the combined model using the ROI_tumor+peritumor 2 demonstrated satisfactory predictive performance. This approach exhibited superior predictive value compared with models relying solely on clinical or radiomic features, thereby facilitating more reliable prediction of the pathological main component in cHCC-CCA.

The combined model developed in this study, which integrates ultrasound radiomics features with clinically relevant indicators, enables non-invasive and effective preoperative prediction of the primary pathological component in cHCC-CCA. However, despite these significant findings, several notable limitations exist. Firstly, although the combined model showed statistically significant differences in AUC values compared with other models in the training set and the overall set, no significant difference was observed in the test set. This may be attributed to the limited sample size enrolled in the present study, indicating that further studies with larger sample sizes are required to validate these findings. Secondly, the study samples were obtained from a single center, and the analysis was conducted solely on two-dimensional ultrasound images. Due to sample size constraints, further exploration of contrast-enhanced ultrasound imaging was not feasible. Thirdly, for patients with multiple intrahepatic lesions, only the lesion with the maximum tumor diameter was enrolled for analysis. Although this is a commonly used strategy at present, it cannot fully take inter-tumoral heterogeneity into consideration in patients with multiple lesions, and this limitation remains to be further explored in future research. Nevertheless, the proportion of patients with multiple liver lesions showed no statistically significant difference between the two groups in our study. This minimized the impact of lesion number on the predictive model and improved its robustness. Additionally, our cohort was retrospectively assembled from histopathologically confirmed cHCC-CCA cases, which may introduce selection bias. A prospective study should be conducted when conditions permit further validation of the reliability of the nomogram. Finally, partial loss of follow-up data prevented further assessment of the relationship between the combined model and patient prognosis.

Conclusion

In conclusion, we have established an integrated model that combines ultrasound-based radiomic features with clinical data. This model demonstrates potential as a reliable, non-invasive method for the preoperative evaluation of the primary component in patients with cHCC-CCA.

Authors’ contributions

RZG, YH and HY designed and directed the study. YPY, WML, YQW and JBP were responsible for the data analysis. YPY, XMB, RW and JSP participated in the writing of the manuscript. All authors were involved in the revision of the manuscript.

Funding

This study was supported by Guangxi Medical University Undergraduate Innovation and Entrepreneurship Training Program (202410598008) and the Clinical Research Climbing Project of The First Affiliated Hospital of Guangxi Medical University (No.YYZS2020024).

Data availability

The dataset and materials for this study can be obtained from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University and conducted in accordance with the Declaration of Helsinki. The requirement of informed consent was waived by the Ethics Committee of First Affiliated Hospital of Guangxi Medical University owing to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.