Application of multimodal fusion technology in early recurrence prediction and pathological analysis of hepatocellular carcinoma

Abstract

BACKGROUND

Early recurrence is an important factor affecting the prognosis of hepatocellular carcinoma (HCC), but its preoperative prediction remains challenging.

AIM

To explore the value of a multimodal interpretable fusion model combining computed tomography (CT) habitat imaging (HI), radiomics, and clinical features in predicting early recurrence of HCC and analyze its correlation with pathological indicators.

METHODS

The 191 HCC patients were categorized into early recurrence and non-early recurrence groups based on postoperative follow-up outcomes, and randomly divided into training and testing sets in a 7:3 ratio. Based on CT arterial phase and clinical data, the habitat model, radiomics model, clinical model, and fusion model were constructed and compared for their predictive ability in early recurrence of HCC. For the optimal model, SHapley Additive exPlanations (SHAP) analysis was performed to evaluate the contribution of different features in the model, and the correlation between HI and radiomics features with tumor microvascular invasion (MVI), Ki67 expression, GPC-3 expression, and pathological grading was analyzed.

RESULTS

The fusion model demonstrated the best performance in predicting early recurrence of HCC, achieving the area under the curve of 0.933 on the validation set. The decision curve analysis curve indicated that the fusion model yielded the highest clinical net benefit. SHAP analysis provided valuable insights into explaining the fusion model's prediction of early HCC recurrence. Correlation analysis revealed significant associations between certain radiomics and Habitat features and pathological indicators such as MVI and Ki-67 expression in HCC.

CONCLUSION

An interpretable fusion model integrating clinical, radiomic, and habitat features can assist clinicians in identifying early postoperative recurrence of HCC, offering significant potential for prognosis prediction and clinical management.

Key Words: Hepatocellular carcinoma; Habitat imaging; Radiomics; Early recurrence; Fusion model

Core Tip: This study explores an interpretable fusion model that combines clinical, radiomics, and habitat features based on enhanced arterial-phase computed tomography images of the liver to predict early recurrence after hepatocellular carcinoma (HCC) surgery. The model outperforms traditional radiomics, clinical, and habitat models. It can help clinicians identify early recurrence of HCC after surgery.

INTRODUCTION

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related deaths globally. Its insidious onset and rapid progression result in most patients being diagnosed at an advanced stage, leading to a poor prognosis[1]. Although radical surgical resection is the main curative treatment for early-stage HCC, the early recurrence rate within two years post-surgery remains as high as 30%-50%, which becomes a critical bottleneck limiting long-term survival[2,3]. Current clinical prognostic indicators, such as alpha-fetoprotein (AFP) and tumor size, can only reflect partial biological characteristics of the tumor and are limited in capturing tumor heterogeneity, reducing the accuracy and clinical utility of preoperative early recurrence risk prediction[4,5]. Therefore, developing imaging tools capable of non-invasive, precise assessment of HCC recurrence risk holds significant clinical importance for optimizing postoperative follow-up strategies and guiding individualized interventions.

Imageomics technology offers a novel approach to quantifying tumor biological behavior by extracting high-throughput texture, morphological, and functional features from medical images, demonstrating potential in HCC diagnosis and prognostic assessment[6,7]. However, traditional radiomics models often treat the tumor as a homogeneous entity for feature extraction, ignoring the tumor's internal heterogeneity (ITH), such as spatial distribution differences in cell density, blood perfusion, and necrosis degree. This significantly reduces the model's accuracy in predicting recurrence risk and limits its clinical translation[7,8]. Habitat imaging (HI), an emerging imaging analysis technology, uses clustering algorithms like K-means to segment the tumor into subregions with different biological characteristics[9-11]. This allows for intuitive and non-invasive quantification of ITH, providing an imaging perspective closer to the pathological essence for revealing malignant phenotypes such as tumor invasiveness and proliferative activity. Studies have confirmed that HI features are superior to traditional radiomics features in prognosis evaluation of malignant tumors such as breast cancer and gliomas. However, research on constructing integrated models using computed tomography (CT)-enhanced arterial-phase HI features combined with clinical and radiomics features for early recurrence prediction of HCC is still limited, and the mechanisms linking these features to key pathological indicators such as microvascular invasion (MVI) and Ki-67 expression are not yet clear.

This study aims to construct a multimodal fusion model based on CT-enhanced arterial-phase images, integrating clinical features, traditional radiomics features, and HI features, and compare its performance with that of single-modality models (clinical model, radiomics model, HI model) to verify its advantages in predicting early postoperative recurrence of HCC. At the same time, SHapley Additive exPlanations (SHAP) analysis will be used to analyze the contribution of key predictive features in the model and explore the correlation between HI and radiomics features and MVI, Ki-67, GPC-3 expression, and pathological grading. The aim is to provide new imaging biomarkers for non-invasive assessment of early recurrence of HCC and promote the application of radiology in precision prognostic management of HCC.

MATERIALS AND METHODS

This study was approved by the Ethics Committee of the Affiliated Hospital of Youjiang Medical University for Nationalities (approval No. YYFY-LL-2024-038). In accordance with the Helsinki Declaration, retrospective studies are exempt from the requirement to obtain informed consent from participants.

Patients

This retrospective study included 191 patients with HCC who were treated at the Affiliated Hospital of Youjiang Medical University for Nationalities from January 2019 to August 2023. Based on postoperative follow-up data, the study subjects were first categorized into an early recurrence group (n = 79) and a non-early recurrence group (n = 112). Subsequently, using random allocation, all patients were further divided into a training set (n = 133) and a test set (n = 58) at a ratio of 7:3.

Inclusion criteria: (1) Patients undergoing radical resection with postoperative pathological confirmation of HCC, possessing comprehensive clinical and follow-up data; and (2) Within one week prior to surgery, perform a contrast-enhanced CT scan of the liver concurrently with the corresponding laboratory tests.

Exclusion criteria: (1) CT images with artifacts during the arterial phase; and (2) Patients with prior antitumor treatments (radiotherapy, chemotherapy, interventional therapy) or concomitant malignancies in other sites.

All patients underwent follow-up at 3-month intervals for the first 2 years post-surgery, then every 6 months thereafter. Each follow-up included enhanced upper abdominal CT or magnetic resonance imaging (MRI) and tumor marker testing.

Clinical features

Clinical and laboratory parameters included age, gender, body mass index, alcohol consumption history, hepatitis B surface antigen, AFP, albumin, aspartate aminotransferase, and alanine aminotransferase.

Pathological indicators

This study referred to previous findings and included four pathological indicators closely related to postoperative recurrence of HCC: MVI[12], Ki-67[13], GPC-3[14], and pathological grading[15], in order to enhance the interpretability of the results. Among these, MVI was categorized into three levels: M0, M1, and M2. Low expression of Ki-67 (< 10%-20%) suggested slow tumor proliferation, better differentiation, and a favorable prognosis, while high expression (≥ 30%-50%) indicated active proliferation, poor differentiation, strong invasiveness, and a higher risk of recurrence[16]. GPC-3 was considered positive when tumor cells showed brown-yellow granular staining in the cytoplasm/membrane, with positive cells ≥ 10%, and negative otherwise[17]. Pathological grading used the Edmondson-Steiner system[18], with differentiation gradually decreasing from grade I to grade IV, accompanied by increasing cellular atypia, nuclear-to-cytoplasm ratio, and a shift in arrangement from regular trabecular to loose and disordered.

Diagnosis of early recurrence of HCC

During the follow-up within 2 years after surgery for HCC patients, the discovery of new tumors in the liver or extrahepatic regions through imaging studies like ultrasound, CT, or MRI can diagnose early recurrence[18]. The diagnosis of tumor recurrence requires joint evaluation by a hepatobiliary surgeon and a radiologist. During the evaluation, both doctors rely solely on the designated clinical data and imaging findings, without access to any irrelevant information. Based on follow-up outcomes: Cases experiencing tumor recurrence within 2 years post-surgery were classified into the recurrence group; those without recurrence were assigned to the non-recurrence group.

CT scanning method

The CT scanning method involved a scanning range from the diaphragm to the inferior margin of the liver. Contrast-enhanced scanning used a high-pressure injector with an 18G intravenous catheter to inject non-ionic iodinated contrast agent at 1.5-2.0 mL/kg at a rate of 3.0-3.5 mL/s, followed by immediate injection of 20-50 mL of saline to flush the catheter. After the bolus injection of the contrast agent, the aortic concentration reached the threshold [120 Hounsfield unit (HU)], triggering the arterial phase scan automatically after a 10-second delay (Table 1 showed the scanning parameters for each device).

Table 1 Scanning parameters of different computed tomography devices.

Devices	Revolution aca (GE)	Ingenuity core 64 (Philips)	Revolution (GE)
Layer thickness (mm)	5	5	5
Layer interval (mm)	5	5	5
Tube voltage (kV)	120	120	120
Tube current (mA)	50	30	50
Matrix	512 × 512	512 × 512	512 × 512
Threshold of ROI (HU)	100	150	120

HU: Hounsfield unit; ROI: Region of interest.

CT image segmentation

Download all patients' liver CT-enhanced arterial phase images in transverse DICOM format and convert them to NIFTI format for storage. Preprocessing of CT image datasets: First, calibrate the voxel spacing of all images to 1 mm × 1 mm × 1 mm. Then, standardize the window width and window level to 259 HU and 40 HU, respectively. Using ITK-SNAP 3.8.0 software, a radiologist with over 10 years of experience independently performed manual slice-by-slice segmentation of tumor regions in the CT arterial phase images (Figure 1). These segmentations were then cross-checked one-by-one by a senior-level physician (the study workflow is illustrated in Figure 2).

Figure 1 Study population and workflow. HCC: Hepatocellular carcinoma; CT: Computed tomography.

Figure 2 Tumor segmentation of a typical case. A: The original computed tomography arterial phase image; B: The tumor segmentation using ITK-SNAP software; C: Pseudo-color image of tumor habitat characteristics.

Radiomics analysis

Radiomics analysis was performed using Pyradiomics (version 3.0.1) to extract 1834 radiomic features from the CT-enhanced arterial phase images, including tumor shape features, first-order features, texture features, and wavelet features. The texture features were derived from the Gray Level Co-occurrence Matrix, Gray Level Size Zone Matrix, Gray Level Run Length Matrix, Neighbouring Gray Tone Difference Matrix, and Gray Level Dependence Matrix.

Habitat analysis

The K-means algorithm, one of the most widely used methods, worked by iteratively finding the result that minimized the sum of the squared distances between each data point and its nearest K cluster centers, dividing the data points into different clusters such that the similarity within the same cluster was high, and the similarity between different clusters was low. Due to its high computational efficiency and strong interpretability, it was particularly suitable for high-dimensional and large-scale data[19]. In this study, we used the K-means clustering algorithm from the scikit-learn library, performing clustering analysis on the voxel signal intensities within the tumor based on arterial phase liver CT enhancement images. The number of clusters (k) ranged from 2 to 5, and through optimization using the Calinski-Harabasz index, k = 3 was determined as the optimal number of clusters. After clustering, the PyRadiomics module was used to extract features from each sub-region, and by calculating the mean of the sub-region features, 5502 representative features (1834 × 3) were obtained.

Feature selection and model construction

Feature selection and model construction involved selecting features with good reproducibility and low redundancy by first applying Z-score normalization to the training and testing set features. Significant features were selected using t-tests or Mann-Whitney U tests (P < 0.05), and distinctive features with a Pearson correlation coefficient > 0.9 were selectively excluded to reduce multicollinearity. The least absolute shrinkage and selection operator (LASSO) algorithm was then used to construct a penalty function λ, shrinking some regression coefficients to zero to select stable features, with the optimal λ value determined by 10-fold cross-validation using the scikit-learn library, completing the feature optimization. Based on the selected features, clinical, radiomics, habitat radiomics, and fusion models were constructed. Model construction was carried out using three machine learning algorithms: RandomForest, ExtraTrees, and multi-layer perceptron (MLP), with the models constructed from the corresponding optimal features being defined as target models. The synthetic minority oversampling technique was combined with oversampling to address the class imbalance in the training set, while the validation set maintained the original sample distribution. The model's robustness is ensured through hyperparameter tuning via 5-fold cross-validation combined with grid search.

Model explanation

Model explanation: In this study, the SHAP package in Python was used for an in-depth analysis of the fusion model, quantifying the importance of each feature to visualize the model's decision logic. SHAP calculates the importance value for each feature and intuitively presents its impact on the model's output in numerical form[19]. Specifically, features with higher average absolute SHAP values play a more crucial role in the model's prediction process. Additionally, using the SHAP summary plot, the degree of feature impact can be presented visually, helping researchers quickly understand the relationship between features and prediction results, providing scientific evidence for model optimization and feature selection.

Statistical analysis

This study conducted data processing and model construction in a Python 3.7.12 environment: Statistical analysis was performed using Statsmodels 0.13.2, while machine learning algorithms were implemented with Scikit-learn 1.0.2. Continuous variables underwent intergroup comparisons via t-tests or Mann-Whitney U tests based on normality. Model diagnostic performance was evaluated using receiver operating characteristic curves, with clinical utility validated through decision curve analysis. For the model with optimal performance, Pearson correlation analysis was performed in R software (version 4.2.2, http://www.Rproject.org) to investigate the association among the selected features and MVI, Ki67 expression, GPC-3 expression, and pathological grade in HCC.

RESULTS

Analysis of clinical features

The patients' baseline clinical characteristics are shown in Table 2. Training set analysis revealed significant differences in AFP levels (P = 0.012) and gender (P = 0.016) between the group with early HCC recurrence and the group without early recurrence. No statistically significant differences were observed for other clinical indicators between the two groups (P > 0.05). Therefore, AFP and sex were selected for clinical model building.

Table 2 Baseline clinical characteristics analysis.

Clinical feature	Training set (n = 133)			Validation set (n = 58)
	No early recurrence (n = 77)	Early recurrence (n = 56)	P value	No early recurrence (n = 35)	Early recurrence (n = 23)	P value
Age (year)	54.49 ± 10.24	51.04 ± 10.88	0.063	50.46 ± 10.31	50.17 ± 9.73	0.917
BMI (kg/m2)	22.26 ± 3.46	22.00 ± 2.54	0.632	21.71 ± 3.41	22.80 ± 2.29	0.183
HBsAg (IU/mL)	967.49 ± 869.13	1019.11 ± 726.57	0.496	1141.26 ± 1046.98	1141.35 ± 762.88	0.893
AFP (ng/mL)	369.27 ± 494.96	539.03 ± 519.80	0.012	523.78 ± 532.39	473.20 ± 539.57	0.975
ALB (g/L)	39.35 ± 5.21	39.18 ± 4.60	0.842	41.55 ± 7.36	38.83 ± 4.59	0.069
AST (IU/L)	82.31 ± 135.98	51.94 ± 27.49	0.758	64.54 ± 69.10	68.47 ± 71.76	0.169
ALT (IU/L)	75.34 ± 118.98	50.77 ± 37.14	0.531	62.21 ± 77.61	74.23 ± 98.27	0.474
Sex, n (%)			0.016			0.08
Female	17 (22.08)	3 (5.36)		9 (25.71)	1 (4.35)
Male	60 (77.92)	53 (94.64)		26 (74.29)	22 (95.65)
Drinking, n (%)			0.173			1
No	54 (70.13)	32 (57.14)		24 (68.57)	15 (65.22)
Yes	23 (29.87)	24 (42.86)		11 (31.43)	8 (34.78)

BMI: Body mass index; HBsAg: Hepatitis B surface antigen; AFP: Alpha-fetoprotein; ALB: Albumin; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase.

In terms of physical characteristics, the average Ki-67 expression index among 191 HCC patients was 0.23 (standard deviation 0.21); the MVI incidence rate was 35.08% (67/191). Among the MVI-positive patients, 37 were in the low-risk group (M1), and 30 were in the high-risk group (M2). The pathological grade distribution was as follows: 13 cases (6.81%) at Edmondson-Steiner grade I, 82 cases (42.93%) at grade II, 73 cases (38.22%) at grade III, and 23 cases (12.04%) at grade IV. GPC-3 expression was positive in 163 cases and negative in 28 cases. Table 2 provides more detailed information about all the patients.

Among the three models constructed with AFP and gender, RandomForest, ExtraTrees, and MLP, the RandomForest model demonstrated the best overall predictive performance, with test set accuracy, area under the curve (AUC), sensitivity, and specificity of 0.517, 0.650, 0.870, and 0.286, respectively. Therefore, it was chosen as the clinical model (Table 3 and Figure 3).

Figure 3 Receiver operating characteristic and decision curve analysis curves for clinical features. A and B: The pink, dark blue, and cyan colors represent the RandomForest, ExtraTrees, and multi-layer perceptron models in the training and validation sets. AUC: Area under the curve; MLP: Multi-layer perceptron.

Table 3 Clinical model construction.

Dataset	Model name	Accuracy	AUC	95%CI	Sensitivity	Specificity
Training	RandomForest	0.677	0.766	0.6884-0.8443	0.893	0.519
Test	RandomForest	0.517	0.650	0.5106-0.7900	0.870	0.286
Training	ExtraTrees	0.632	0.678	0.5881-0.7671	0.661	0.61
Test	ExtraTrees	0.534	0.598	0.4506-0.7457	0.696	0.429
Training	MLP	0.579	0.634	0.5398-0.7281	0.839	0.39
Test	MLP	0.466	0.504	0.3482-0.6605	0.913	0.171

AUC: Area under the curve; MLP: Multi-layer perceptron.

Comparison of prediction results and performance across different models

After performing t-test, Mann-Whitney U test, Pearson correlation analysis, and LASSO regression for dimensionality reduction, 9 radiomics features were selected to construct the radiomics prediction model, 10 habitat features were used to build the habitat prediction model, and 12 fusion features were used to construct the fusion model. The predictive metrics of different models were shown in Table 4 and Figure 4, and the results indicated that the fusion model had the best overall predictive performance. The Delong test (Figure 4) showed that in the test set, the fusion model's AUC (0.933) was significantly better than that of the clinical model (0.650) and radiomics model (0.743) (all P < 0.05); compared with the habitat model (0.830), the fusion model's AUC did not show a statistically significant difference in both the training and test sets (P > 0.05), but the decision curve analysis (Figure 4) indicated that the fusion model had higher clinical net benefit.

Figure 4 Receiver operating characteristic curves and decision curve analysis for subjects with different characteristics. A and B: Receiver operating characteristic curves for the training and testing sets, where yellow, red, cyan, and dark blue represent fusion features, habitat features, radiomics features, and clinical features, respectively; C and D: Delong test analysis for the training and testing sets; E and F: Decision curve analysis for the training and testing sets. Red, blue, green, and yellow represent the integrated model, habitat model, traditional radiomics model, and clinical model, respectively. AUC: Area under the curve; DCA: Decision curve analysis.

Table 4 Predictive indicators for different models.

Dataset	Model	Accuracy	AUC	95%CI	Sensitivity	Specificity
Training	Clinical	0.677	0.766	0.6884-0.8443	0.893	0.519
Test	Clinical	0.517	0.650	0.5106-0.7900	0.870	0.286
Training	Radiomics	0.729	0.785	0.7083-0.8608	0.625	0.805
Test	Radiomics	0.759	0.743	0.6018-0.8840	0.652	0.829
Training	Habitat	0.774	0.887	0.8328-0.9413	0.732	0.805
Test	Habitat	0.741	0.830	0.7239-0.9357	0.87	0.657
Training	Combined	0.857	0.941	0.9058-0.9764	0.946	0.792
Test	Combined	0.845	0.933	0.8735-0.9923	0.826	0.857

AUC: Area under the curve.

SHAP analysis

Further analyze the contribution levels and directional effects of the features within the best-performing ensemble model using SHAP values. The results indicated that h2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis, gender, h3_gradient_firstorder_Skewness, h1_lbp_3D_m1_glszm_HighGrayLevelZoneEmphasis, and h3_wavelet_HHL_firstorder_Kurtosis had a positive impact on predicting short-term recurrence of liver cancer, while h1_original_glszm_ZoneVariance, wavelet_HLH_firstorder_Skewness_CT, and h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis had a negative effect. Among them, h2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis contributed the most to the prediction model (Figure 5), and the other four features had a positive effect. For the first sample, the waterfall plot (Figure 5) and force plot (Figure 5) visualizations show that h2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis, h1_original_glszm_ZoneVariance, gender, h3_gradient_firstorder_Skewness, h1_lbp_3D_m1_glszm_HighGrayLevelZoneEmphasis, h1_log_sigma_3_0_mm_3D_firstorder_90Percentile, and h3_wavelet_HHL_firstorder_Kurtosis played a positive regulatory role in the prediction of this sample, while wavelet_HLH_firstorder_Skewness_CT and h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis played a negative regulatory role. The final predicted probability for this sample was 0.669, indicating a higher risk of short-term recurrence.

Figure 5 SHapley Additive exPlanations analysis plot. A: SHapley Additive exPlanations adds feature importance explanations, represented by the average absolute shape value for each feature; B: The shape summary plot displays the distribution of shape values for each feature across patients. Each point corresponds to the shape value of a specific feature for each patient. The X-axis represents shape values, with colors ranging from red (high feature value) to blue (low feature value); C: Shape plots for patients with and without recurrence. Colors indicate each feature's contribution (red for positive, blue for negative), while bar length reflects contribution strength. SHAP: SHapley Additive exPlanations; CT: Computed tomography.

Correlation analysis

To further explore the potential biological significance of imagingomics/habitat features, we conducted an in-depth analysis of the correlations between the selected features and MVI, Ki-67 exposure, GPC-3 expression, and tumor grade. The results are presented in Table 5 and Figure 6. Correlation analysis showed that wavelet_HLH_firstorder_Skewness_CT and h3_wavelet_HHL_firstorder_RootMeanhSquared were significantly positively correlated with GPC-3 expression, h1_log_sigma_3_0_mm_3D_firstorder_90Percentile was significantly positively correlated with MVI and Ki-67 expression, h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis was significantly negatively correlated with both MVI and Ki-67 expression, while h1_lbp_3D_m1_glszm_HighGrayLevelZoneEmphasis was significantly positively correlated with MVI expression.

Figure 6 Correlation heatmap. The numbers and colors within the squares indicate the correlation coefficients between variables, where warm colors (e.g., red) represent negative correlations and cool colors (e.g., blue) represent positive correlations. The deeper the color, the stronger the correlation. ^aP < 0.05; ^bP < 0.01; ^cP < 0.001. MVI: Microvascular invasion.

Table 5 Correlation analysis between different characteristics and microvascular invasion, Ki67, GPC-3 expression, and pathological grading in hepatocellular carcinoma.

	MVI		Ki-67		GPC-3		Grading
	r value1	P value	r value1	P value	r value1	P value	r value1	P value
wavelet_HLH_firstorder_Skewness_CT	0.0535	0.4625	-0.0461	0.5267	0.1425	0.0492	0.0441	0.5451
h2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis	-0.0148	0.8389	0.0354	0.6271	0.0042	0.9535	-0.0664	0.3613
h1_log_sigma_3_0_mm_3D_firstorder_90Percentile	0.1792	0.0131	0.1462	0.0435	0.0572	0.4319	-0.0196	0.7883
h3_wavelet_HLH_gldm_SmallDependenceHighGrayLevelEmphasis	0.0758	0.2975	0.0514	0.4805	0.0077	0.9154	-0.0605	0.4056
h3_wavelet_HHL_firstorder_RootMeanh3_squared	0.0687	0.3452	-0.0729	0.3163	0.1703	0.0185	0.0198	0.786
h3_wavelet_HHL_firstorder_Kurtosis	0.028	0.7007	-0.0464	0.5238	-0.0733	0.3133	0.001	0.9891
h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis	-0.3128	< 0.001	-0.1574	0.0297	-0.1179	0.1042	0.0141	0.8467
h1_logarithm_ngtdm_Coarseness	-0.0491	0.5004	-0.0097	0.8940	0.0301	0.6797	0.1327	0.0672
h1_original_glszm_ZoneVariance	0.0926	0.2024	0.0223	0.7596	0.0289	0.6918	0.1272	0.0796
h3_gradient_firstorder_Skewness	0.1075	0.1389	0.0235	0.7474	0.113	0.1195	0.0261	0.7201
h1_lbp_3D_m1_glszm_HighGrayLevelZoneEmphasis	0.159	0.0281	0.1189	0.1013	0.0419	0.5651	0.0321	0.6594

¹Pearson correlation coefficient.

DISCUSSION

In this study, we developed a fusion model based on CT arterial phase images, alongside habitat models, radiomics models, and clinical models. We compared their predictive capabilities for early recurrence of HCC and assessed the relevance of screened features to lesion parameters. A results show that the fusion model demonstrated the highest predictive performance. The results indicated that the fusion model had the best predictive performance. Some of the fused features showed significant correlations with MVI, Ki-67, and GPC-3 expression in HCC. The study results demonstrate strong innovation and clinical value.

The fusion model demonstrated significant advantages in predicting early recurrence of HCC, achieving a verification set AUC of 0.933. This performance markedly surpassed that of the habitat model (AUC = 0.830), traditional radiomics model (AUC = 0.743), and clinical model (AUC = 0.650). This advantage is similar to previous results of fusion models based on CT or MRI[20-23], further confirming the scientific validity of "multi-feature fusion enhancing prediction accuracy". At the same time, decision curve analysis shows that it provides greater net clinical benefit at various thresholds, which together confirm the clinical utility of the fusion model, significantly surpassing other single-type models. This advantage may stem from multi-dimensional support: (1) In terms of feature comprehensiveness, the fusion model does not rely on a single information source but innovatively integrates clinical features (such as gender, AFP levels), traditional radiomics features (such as gray-level co-occurrence matrix, texture statistical parameters), and HI features (such as tumor sub-region blood flow, cell density distribution). Compared to clinical models that rely solely on limited clinical indicators and cannot reflect tumor heterogeneity[24], and traditional radiomics models that depend on manually designed texture features and are susceptible to subjective bias and functional redundancy[24,25], the fusion model, through multi-modal feature complementarity, achieves a "panoramic" depiction of tumor biological behaviors (such as invasiveness, proliferative activity) and individual patient differences, capturing richer recurrence-related information, laying the foundation for high-precision prediction; and (2) The unique value of habitat features adds to the advantage. The habitat model quantifies tumor heterogeneity non-invasively by dividing the tumor into sub-regions with specific physiological states (such as blood flow, necrosis, and edema)[11], and tumor heterogeneity is a core driver of HCC invasion and recurrence[26]. This is also the key reason why the habitat model (AUC = 0.830) outperforms traditional radiomics and clinical models. The fusion model further integrates habitat features with other features, essentially adding clinical risk stratification, microscopic texture features, and other dimensions on top of "heterogeneity assessment", which further amplifies the ability to identify recurrence risk, breaking the limitations of single-model predictions.

SHAP analysis is an important method for understanding the decision-making mechanism of ensemble models. The results show that the five features: H2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis, gender, h3_gradient_firstorder_Skewness, h1_lbp_3D_m1_glszm_HighGrayLevelZoneEmphasis, and h3_wavelet_HHL_firstorder_Kurtosis have a positive effect on early recurrence of HCC. The three features h1_original_glszm_ZoneVariance, wavelet_HLH_firstorder_Skewness_CT, and h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis are negatively correlated with the risk of short-term recurrence. Among them, h2_wavelet_LHH_glszm_LargeAreaLowGrayLevelEmphasis has the highest contribution to the model, indicating that different fusion features play different roles in the model's decision-making. Positive features may be associated with malignant tumor phenotypes (such as invasiveness and proliferative activity), such as MVI and tumor volume size[27,28], and their high expression or abnormalities may promote short-term recurrence of liver cancer. Negative features may be related to tumor suppression or good prognosis phenotypes, such as higher ADC values and intact capsules[29], and their high expression may reduce the risk of short-term recurrence. Waterfall and force plot visualizations can intuitively identify the contribution and direction of each feature, providing strong support for clinicians to understand the model's prediction results, which helps improve the model's explainability and clinical confidence[30].

The relevance of features to disease mechanisms provides the fusion model with superior biological rationale. Existing research indicates that MVI serves as a crucial pathological biomarker for assessing tumor aggressiveness and migratory potencyl[28], and the expression level of KI-67 is closely associated with tumor cell proliferation activity[31]. Furthermore, HCC patients with higher MVI grades and Ki-67 expression indices have a significantly increased risk of postoperative recurrence[12,32]. This study found that some fusion features are directly associated with key pathological parameters of HCC: For example, h1_log_sigma_3_0_mm_3D_firstorder_90Percentile shows a significant positive correlation with MVI and Ki-67 expression, while h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis shows a significant negative correlation with both MVI and Ki-67 expression. This suggests that HCC with high h1_log_sigma_3_0_mm_3D_firstorder_90Percentile values and low h1_wavelet_LHH_glrlm_LowGrayLevelRunEmphasis values have higher aggressiveness and tumor cell active multiplication, making these patients more likely to experience short-term recurrence. This aligns with the impact direction of these characteristics on the model as shown in SHAP analysis. wavelet_HLH_firstorder_Skewness_CT (r = 0.1425, P = 0.0492) and h3_wavelet_HHL_firstorder_RootMeanh3_squared (r = 0.1703, P = 0.0185) are associated with GPC-3 positive expression. GPC-3, as a marker for liver cancer stem cells, exhibits high expression indicating a dedifferentiated state of tumor cells[33]. Both these characteristics reflect irregular gray-level distribution within tumors, consistent with the microenvironmental features of stem cell nests. This suggests that these two features hold promise as non-invasive imaging markers for indirectly assessing stem cell-related malignant phenotypes (such as dedifferentiation levels) within tumors of liver cancer patients. Of course, the GPC-3 correlation analysis in this study is exploratory in nature. The effect size of this association is limited, and the sample size is relatively small. The robustness of the weak correlation results still requires validation with larger samples. This implies that the fusion model's predictions are not solely based on “data fitting”, but rather reflect the imaging manifestation of tumor molecular biological processes through extracted features. Consequently, it establishes an intrinsic connection with pathological indicators, rendering its predictions more pathologically interpretable and better equipped to guide clinical assessments of tumor essence.

In conclusion, the fusion model demonstrates significant advantages in predicting early recurrence of HCC, due to the comprehensiveness of feature dimensions, the uniqueness of habitat features, and the correlation with pathological mechanisms. It not only provides a high-precision prediction tool for clinical use but also offers a new direction for HCC recurrence research and individualized treatment through the association of "imaging features - pathological mechanisms". However, this study also has several limitations: (1) This study is a retrospective, single-center investigation lacking external validation. Furthermore, the lack of statistically significant improvement in the AUC of the fusion model may be attributed to insufficient statistical power due to the small sample size. Future research should expand the sample size and incorporate data from multicenter studies to enhance the generalizability and reliability of the findings; (2) The research process relies on the physician's experience and skill. When manually delineating lesions, subjective bias is unavoidable. In the future, a combined manual and semi-automated delineation approach will be adopted, where physicians only refine blurred boundaries. This aims to minimize subjective bias and enhance the accuracy of lesion delineation; (3) When extracting ITH using HI, this study exclusively utilized arterial phase images from contrast-enhanced CT scans, whose precision remains significantly limited. Future efforts will optimize the HI algorithm by integrating multimodal imaging to enhance ITH extraction accuracy. Large-scale studies will correlate HI habitat maps with histopathological and genomic data to elucidate the biological significance of each subregion, providing theoretical support for understanding tumor characteristics; and (4) This study did not concurrently collect pathological slide data (e.g., tumor necrosis fraction, microvascular density measurements) from enrolled patients. It relied solely on imaging features to infer correlations with pathological indicators, lacking direct validation. Future work will integrate pathological slides to more thoroughly explore the relationship between “imaging features and pathological mechanisms”.

CONCLUSION

This study, based on CT arterial-phase images and combining clinical features, radiomics features, and HI characteristics, constructed a fusion model that has significant advantages in predicting early recurrence of HCC. Through SHAP analysis and correlation studies, the model's decision logic and the biological significance of key features were revealed. Despite its limitations, this multi-feature fusion framework provides an innovative approach for early recurrence assessment of HCC. It lays the foundation for clinical translation and has the potential to advance the application of imaging in precise diagnosis and treatment of HCC. It is expected to help improve the level of patient prognosis management.

ACKNOWLEDGEMENTS

We would like acknowledge all funding organizations and patients of this study.