A CECT‐Based 2PI System as a Novel Noninvasive Prognostic Tool for Hepatocellular Carcinoma: A Dual‐Validation Study

ABSTRACT

This study aimed to develop and validate machine learning (ML) models integrating clinical parameters and the 2PI system (Pathology and Prognosis‐Informed Imaging System) for predicting postoperative recurrence risk in hepatocellular carcinoma (HCC). The multicenter retrospective study included 496 patients with solitary HCC (≤ 5 cm). Surgical resection (SR) patients from the primary center constituted the training set; radiofrequency ablation (RFA) patients from the same center formed the internal test set; and SR patients from other centers served as the external test set. In the training set, multivariable logistic regression identified seven imaging features associated with pathological markers, with odds ratios calculated as exp(β). Kendall's tau‐b coefficient was used to indicate the strength of association between pathology and recurrence risk. A dual‐information pathway 2PI system was constructed based on these coefficients, and the performance of ML models integrating clinical parameters and the 2PI system was evaluated. A total of 496 patients (mean age 58 ± 10 years, 376 men) were enrolled. Patients were stratified into high‐ and low‐risk groups using a 2PI system threshold of ≥ 19. The Random Survival Forest (RSF) model incorporating clinical parameters and the 2PI system demonstrated favorable predictive performance across all sets (training: C‐index 0.76 [95% CI: 0.72–0.80]; internal test: C‐index 0.69 [95% CI: 0.63–0.75]; external test: C‐index 0.68 [95% CI: 0.57–0.79]). The 2PI system applies a joint weighting strategy to create a straightforward image scoring system that enhances postoperative recurrence risk prediction for solitary HCC, demonstrating preliminary generalizability across both SR and RFA cohorts.

Using differential weighting, we developed a 2PI‐based risk stratification system for HCC recurrence. A comprehensive prognostic model incorporating clinical variables and the 2PI score further enhanced postoperative prediction in solitary HCC and showed initial generalizability across SR and RFA cohorts.

Abbreviations

2PI system

pathology and prognosis‐informed imaging system

ACLD

advanced chronic liver disease

AFP

alpha‐fetoprotein

APHE

arterial phase hyperenhancement

CECT

contrast‐enhanced computed tomography

C‐index

concordance index

DCA

decision curve analysis

DT

decision tree

HCC

hepatocellular carcinoma

LI‐RADS

liver imaging reporting and data system

LR‐M

LI‐RADS category M

ML

machine learning

MTM

macrotrabecular‐massive

MVI

microvascular invasion

PHT

portal hypertension

RFA

radiofrequency ablation

RFS

recurrence‐free survival

ROC

receiver operating characteristic

RSF

random survival forest

SHAP

SHapley Additive exPlanations

SR

surgical resection

XGBoost

extreme gradient boosting

1. Introduction

The 2020 global cancer statistics showed primary liver cancer as the third leading cause of cancer‐related mortality worldwide, with deaths projected to reach 1.3 million by 2040 (a 56.4% increase compared with 2020 data) [1, 2]. Hepatocellular carcinoma (HCC), accounting for 75%–85% of primary liver cancers, is recognized as an intricate ecosystem defined by pronounced spatiotemporal heterogeneity and has a five‐year survival rate of only about 22% [3, 4]. Although various staging criteria exist, determining the optimal intervention strategy for solitary HCC patients, particularly those with early‐stage tumors and advanced chronic liver disease (ACLD) but preserved liver function (Child‐Pugh A or B), remains challenging [4, 5, 6, 7]. Therefore, a simple, practical, and reliable risk‐prediction tool is urgently needed.

Microvascular invasion (MVI) is a recognized major driver of recurrence in HCC and is closely linked to tumor metastasis, substantially compromising survival outcomes in the HCC population [8]. Emerging evidence indicates that the macrotrabecular‐massive (MTM) subtype, an angiogenesis‐related aggressive histologic phenotype, is likewise a key determinant of poor prognosis following curative treatment [9]. Accumulating data highlight the prognostic significance of pathological and molecular features in HCC [10, 11]. However, the invasiveness of pathological sampling often leads to patient reluctance.

Recent studies have demonstrated that imaging not only reflects macroscopic tumor morphology but also serves as a surrogate for pathological correlates, thereby functioning as a non‐invasive imaging biomarker for HCC [12]. Imaging allows a comprehensive depiction of tumors and their microenvironment [13]. Nevertheless, models focusing solely on tumor‐centric features while disregarding the hepatic background inevitably obscure fundamental factors required for improved risk stratification [14].

Regardless of etiology, cirrhosis and portal hypertension (PHT) constitute major risk factors for HCC [15]. According to the “seed and soil” hypothesis, tumor progression is interactively shaped by the intrinsic properties of the “seed” and the supportive conditions of the “soil”. On the one hand, it is essential to elucidate the intrinsic biological behavior of the MTM subtype as the “seed”; on the other hand, clarifying the pathological state of ACLD with cirrhosis and PHT, the hepatic “soil”, is equally critical. MVI embodies the direct reflection of the interaction between aggressive tumor biology (“seed”) and fragile hepatic substrate (“soil”) [11]. Therefore, integrative evaluation of these three determinants is fundamental to uncovering imaging‐pathology‐prognosis associations and is a prerequisite for establishing imaging biomarkers capable of guiding clinical decision‐making. Notably, conventional contrast‐enhanced computed tomography (CECT), through multiphase imaging, enables assessment of intratumoral and peritumoral features as well as portal venous morphology, thereby reflecting underlying pathological correlates [4, 16].

This study seeks to develop a machine learning (ML) model combining clinical parameters with the 2PI system (Pathology and Prognosis‐Informed Imaging System) to evaluate its incremental prognostic value in predicting HCC recurrence. The significance of this work lies in the ability of the 2PI system to translate imaging features into quantitative metrics through a differential weighting approach. Beyond illuminating the deep associations among imaging signatures, pathological attributes, and recurrence risk, it offers clinicians a practical tool to more readily identify high‐risk patients.

2. Materials and Methods

2.1. Study Population

This multicenter retrospective study was approved by the ethics committees of three tertiary medical centers, with a waiver of informed consent (AAF‐IRB‐029‐06). HCC individuals who received surgical resection (SR) or radiofrequency ablation (RFA) from January 2015 to December 2022 across the three participating centers were initially evaluated for eligibility. Pathological findings from resection specimens served as the reference standard for surgically treated patients, whereas clinical diagnosis combined with follow‐up data was used for those receiving RFA. Inclusion criteria were: (a) patients aged ≥ 18; (b) pathological diagnosis (SR cohort) or clinical diagnosis (RFA cohort) of solitary HCC; (c) radiological evidence of chronic liver disease, manifested by hepatic morphological changes or complications of portal hypertension (slight ascites, splenomegaly, presence of collateral circulation, or main portal vein dilation); (d) availability of complete clinical, follow‐up, and pathological data (pathology applicable to surgical cases only); (e) CECT performed within 1 month before the intervention. Exclusion criteria were: (a) multiple nodules or tumor size ≥ 5 cm; (b) presence of multiple vascular invasions; (c) receipt of any HCC‐related treatment before the initial CECT; (d) CECT image quality system < 3; (e) incomplete clinical, imaging, or follow‐up data.

2.2. CT Study Protocols and Radiologic Assessment

All patients underwent multiphasic CECT. Detailed imaging protocols are provided in Doc S1 and Table S1.

All CECT images were retrospectively interpreted by two radiologists (YZ.C., with 5 years of experience in abdominal radiology, and L.Z., with 20 years of experience in abdominal radiology). Readers were blinded to clinical and pathological information. Eleven imaging features were assessed, including tumor characteristics (non‐rim arterial phase hyperenhancement (APHE), non‐peripheral washout, LI‐RADS category M (LR‐M), intratumoral arteries, nonsmooth tumor margin, incomplete “capsule”, and morphological type), as well as peritumoral findings (wedge‐shaped enhancement, corona enhancement) and portal venous morphological traits (main portal vein diameter and presence of collateral circulation). Feature definitions are detailed in Doc S2. For qualitative variables, discrepancies across readers were resolved by consultation with a senior radiologist, and the consensus results were used for analysis. For measurements of the main portal vein diameter, their mean values were adopted for subsequent statistical analysis.

2.3. CT Image Quality Assessment and Inter‐ and Intra‐Observer Agreement

To guarantee the quality and reliability of the imaging data used in this study, we conducted a prospective quality‐assurance evaluation integrating objective metrics and subjective assessments. For objective quality control, the contrast‐to‐noise ratio (CNR) of each dataset was calculated and compared with the standards in the American College of Radiology (ACR) Quality Control Manual (Table S2). For subjective quality control, all images were scored using a Likert‐scale evaluation (Doc S3).

Imaging features were independently assessed by senior radiologists, with inter‐and intra‐observer consistency measured via Cohen's kappa coefficient. Intra‐observer kappa values were 0.63 to 1.00 (mean: 0.83), and inter‐observer kappa values similarly ranged from 0.63 to 1.00 (mean: 0.85). According to the Landis‐Koch criteria, the mean inter‐ and intra‐observer agreement reached the category of “almost perfect,” providing strong support for the reliability and reproducibility of the evaluative methodology applied in this study. Further details are provided in Figure S1.

2.4. Data Collection

Clinical data, including liver function indices, Alpha‐fetoprotein (AFP), tumor size, sex, and age, were obtained from the initial admission evaluation, performed within 1 month before treatment. Patients undergoing SR received pathological examination following curative hepatectomy. Pathological evaluation was based on hematoxylin–eosin and immunohistochemical staining. Senior pathologists reviewed the slides and extracted the high‐risk pathological features such as MVI, MTM, and cirrhosis. A subset of representative cases had undergone prospective one‐to‐one radiologic‐pathologic correlation in the original study (Doc S4). Follow‐up data for encompassing patients were available; the follow‐up protocol is detailed in Doc S5.

2.5. Establishment and Evaluation of the 2PI System

First, multivariable logistic regression analysis was performed in the training set to identify seven imaging predictors linked to the high‐risk pathological features (MVI, MTM, and cirrhosis), yielding corresponding exp(β) values, which represented the odds ratios (ORs). Second, Kendall's tau‐b coefficients (τ_MVI, τ_MTM, and τ_Cirrhosis) were used to indicate the strength of association between each pathological feature and recurrence risk. Third, seven imaging features were scored and aggregated according to these weighted parameters to construct the 2PI system, an integrated index combining pathological features and prognostic weighting to predict postoperative recurrence risk in HCC. The optimal cutoff was determined via the maximally selected rank statistics approach. Recurrence‐free survival (RFS) was compared across high‐ and low‐risk cohorts among multiple datasets utilizing Kaplan–Meier curves and the log‐rank test.

2.6. Establishment of the HCC Prognostic Model and Assessment of Clinical Utility

Independent HCC recurrence prognostic factors were identified via univariable and multivariable Cox proportional hazards regression analyses, with hazard ratios and 95% CIs derived accordingly. Based on these variables, a clinical prognostic model was established. The performance of the clinical model and the combined model (clinical variables plus 2PI system) was compared using four ML approaches: Cox regression, decision tree (DT), random survival forest (RSF), and extreme gradient boosting (XGBoost) survival analysis. Differences in concordance index (c‐index) were assessed using the two‐sided DeLong test. The clinical utility was confirmed utilizing RFA and SR cohorts. The final model was selected based on an integrated assessment of C‐index and time‐dependent receiver operating characteristic (ROC) curves. The decision curve analysis was utilized to examine clinical applicability. Global model interpretability was achieved via SHapley Additive exPlanations (SHAP).

2.7. Statistical Analysis

2.7.1. Derivation of Feature Scores Based on Weighted Coefficients

Step 1. Multivariable logistic regression analysis was performed to identify imaging features associated with MVI, MTM, and cirrhosis. The derived exp(β) values (ORs) were utilized as the imaging‐pathology association weights for subsequent scoring.

Step 2. Kendall's tau‐b coefficients were employed to quantify the correlation of these pathological features (MVI, MTM, and cirrhosis) with HCC recurrence, yielding respective coefficients.

Step 3. Subsequently, each Kendall's tau‐b coefficient was multiplied by the corresponding parameter's weight (exp(β)). The sum of these products was scaled and rounded to the nearest integer to derive the final score (Doc S6).

2.7.2. General Statistical Methods

Sample‐size calculations are provided in Doc S7. Continuous variables are presented as mean ± standard deviation or median with interquartile range (IQR). Normally distributed continuous variables were compared using Student's t‐test, whereas non‐normally distributed variables were compared through the Mann–Whitney U test. Categorical variables were presented as counts (percentages) and compared using the chi‐square test. All statistical tests were two‐sided, with p < 0.05 considered statistically significant. Statistical analyses were performed using R version 4.4.2.

3. Results

3.1. Baseline Characteristics of the Study Cohort

Among 1436 HCC patients who underwent CECT, 940 were excluded as per the eligibility criteria, leaving 496 eligible patients for analysis (mean age 58 ± 10 years; 376 men) (Figure 1). The training set comprised 317 SR patients from our center (mean age 57 ± 10 years; 247 men). The internal test set encompassed 132 patients treated with RFA at our center (mean age 60 ± 10 years; 94 men). The external test set consisted of 47 patients from two other centers (mean age 59 ± 10 years; 35 men). The mean follow‐up duration was 20 months, during which 235 patients developed recurrence. Demographic and clinical characteristics of the three sets are displayed in Table 1. Subtype analysis results are shown in Figure S2.

FIGURE 1.

Patient selection strategy and cohort assignment.

TABLE 1.

Baseline characteristics of patients.

Characteristics	Training set	Internal test set	External test set	p overall
No. of patients	317	132	47
Age				0.165
60 years	194 (61.2)	70 (53.0)	24 (51.1)
≥ 60 years	123 (38.8)	62 (47.0)	23 (48.9)
Gender				0.311
Male	247 (77.9)	94 (71.2)	35 (74.5)
Female	70 (22.1)	38 (28.8)	12 (25.5)
Etiology				0.363
Hepatitis B	253 (79.8)	112 (85.0)	32 (68.0)
Other a	64 (20.2)	20 (15.0)	15 (32.0)
Laboratory data
ALT (U/L) b	30.3[19.1;46.3]	30.9[20.3;48.8]	27.0[19.9;55.3]	0.812
ALP (U/L) b	79.9[65.6;97.5]	95.4[75.8;124]	71.0[60.5;96.3]	< 0.001
AFP (ng/mL)				0.487
≤ 200	240 (75.7)	104 (78.8)	33 (70.2)
> 200	77 (24.3)	28 (21.2)	14 (29.8)
Child‐Pugh stage	A/B	A/B	A/B	0.516
Tumorsize b	3.20 [2.40;3.90]	2.35 [1.78;3.23]	3.80 [2.50;5.00]	< 0.001
RFS b	20.0 [7.00;37.0]	18.0 [7.75;31.2]	24.0 [9.00;33.5]	0.806

Note: Unless otherwise specified, data are presented as numbers of patients, with percentages in parentheses. p < 0.05 was considered statistically significant.

Abbreviations: AFP, alpha‐fetoprotein; ALP, alkaline phosphatase; ALT, alanine aminotransferase; RFS, recurrence‐free survival.

^a Including other viral hepatitis (excluding hepatitis B), as well as autoimmune hepatitis and drug‐induced liver injury.

^b Data are presented as medians, with IQRs in parentheses.

3.2. Development of the 2PI System

Kaplan–Meier survival analyses first confirmed that the pathological features of MVI, MTM, and cirrhosis were key determinants of patient prognosis (Figures 2A–C). On this basis, a multivariable logistic regression analysis was conducted in the training set, identifying seven imaging features linked to MVI, MTM, and cirrhosis: A1, nonsmooth margin; A2, incomplete “capsule”; A3, intratumoral arteries; A4, corona enhancement; A5, LR‐M category; A6, main portal vein diameter; and A7, collateral circulation. Imaging features stratified by pathological characteristics are presented in Table S3. The exp(β) was first calculated for the seven imaging features (Table 2). Subsequently, the correlation coefficients between the three pathological features and prognosis (τ_MVI = 0.258, τ_MTM = 0.235, τ_Cirrhosis = 0.272) were multiplied by exp(β) of each imaging parameter (Table S4). The scores were calculated by summing these weighted values, scaling the sum by a factor of 10, and rounding to the nearest integer, thereby yielding a composite scoring model. The scoring formula was defined as follows:

FIGURE 2.

Kaplan–Meier analysis of recurrence‐free survival (RFS) stratified by key pathological features and the 2PI system. (A–C) Patients with positive pathological findings had significantly lower RFS rates than those with negative findings (p < 0.05); (D–F) In the training set, internal test set and external test set, patients in the low‐risk group (2PI score < 19) exhibited significantly superior RFS compared with the high‐risk group (score ≥ 19) (p < 0.05). RFS, recurrence‐free survival.

TABLE 2.

Multivariable logistic regression analysis of imaging features associated with pathological characteristics.

Feature ID	Image feature	MVI		MTM subtype		Cirrhosis
		exp(β) a	p	exp(β) a	p	exp(β) a	p
A1	Nonsmooth margin	3.357	< 0.001	0.790	0.004
A2	Incomplete “capsule”	2.756	0.001
A3	Intratumoral arteries	1.075	0.001
A4	Corona enhancement	1.823	< 0.001
A5	LR‐M category			0.990	< 0.001
A6	Main pv diameter					0.729	0.006
A7	Collateral circulation					1.671	< 0.001
Intercept		−6.873		1.357		0.030

Note: p < 0.05 was considered statistically significant.

Abbreviations: LR‐M, LI‐RADS category M; MTM, macrotrabecular‐massive; MVI, microvascular invasion; PV, portal vein.

^a Exp(β) is the natural exponential function of the β value, representing the odds ratio (OR) in our model. Exp(β) is reported directly to reflect the raw output of the model parameters, ensuring transparency and traceability of the results.

The final feature scores of the imaging features were obtained as follows: A1, 11 points; A2, 7 points; A3, 3 points; A4, 5 points; A5, 2 points; A6, 2 points; and A7, 5 points (Table S5). We performed a sensitivity analysis by applying different weighting strategies. The joint weighting strategy outperformed the single‐coefficient method in test sets (Table S6).

3.3. Prognostic Risk Stratification Using the 2PI System

The optimal cutoff point for the 2PI system was 19 using maximally selected rank statistics (Figure 3). A total score ≥ 19 was classified as high risk, whereas scores < 19 were deemed low risk. Across the three datasets, 235 patients (47%) experienced recurrence during the median 20‐month follow‐up. Kaplan–Meier survival analyses demonstrated notable differences between high‐ and low‐risk groups in the SR and RFA cohorts, with high‐risk patients exhibiting consistently poorer RFS (p < 0.001; Figures 2D–F). Furthermore, these clinical prognostic disparities were underpinned by distinct biological characteristics, as the high‐ and low‐risk groups exhibited significant differences in gene expression profiles and immune microenvironment features (Doc S8 and Figure S3). Additionally, we performed exploratory subgroup analyses using Kaplan–Meier curves to observe the prognostic outcomes of RFA and SR within each risk group (Figure S4).

FIGURE 3.

Matching CECT features to high‐risk pathology and visualizing prognosis via 2PI system. (A) Seven imaging features significantly associated with pathological indicators and their corresponding scores. (B) Representative microscopic images (hematoxylin–eosin staining) showing key pathological features. Multivariate logistic regression analysis identified imaging features independently associated with pathological outcomes. MTM subtype was predicted by imaging features A1 (Nonsmooth margin) and A5 (LR‐M); MVI positivity was predicted by A1 (Nonsmooth margin), A2 (Incomplete “capsule”), A3 (Intratumoral arteries), and A4 (Corona enhancement); liver cirrhosis (histopathologically defined as Scheuer stage S4 or nodular cirrhosis) was predicted by A6 (Main pv diameter) and A7 (Collateral circulation). (C) The optimal cutoff value of the 2PI system for predicting HCC recurrence was determined using the maximally selected rank statistics method based on RFS, with scores ≥ 19 defined as high‐risk and < 19 as low‐risk. MTM, macrotrabecular‐massive; MVI, microvascular invasion; pv, portal vein.

3.4. Univariable and Multivariable Cox Regression Analyses

Five variables were associated with recurrence in the univariable Cox regression analyses, and four variables, ALT, ALP, AFP, and tumor size were retained as predictors in the clinical model (Table 3). Multivariable Cox regression of the combined model integrating the clinical factors and the 2PI system further revealed that the 2PI system (HR, 1.05 [95% CI, 1.03–1.08]; p < 0.001), ALP (HR, 1.01 [95% CI, 1.00–1.01]; p = 0.027), and AFP (HR, 1.54 [95% CI, 1.07–2.22]; p = 0.021) emerged as independent HCC recurrence predictors (Table S7).

TABLE 3.

Univariate and multivariate Cox regression analyses for model building.

Characteristic	Univariable analysis		Multivariable analysis
	Hazard ratio	p*	Hazard ratio	p*
Age	1.01 (0.99–1.03)	0.208
Gender	0.91 (0.61–1.35)	0.640
Size	1.20 (1.03–1.40)	0.020	1.18 (1.01–1.38)	0.037
AST	1.00 (1.00–1.01)	0.020	0.99 (0.99–1.00)	0.122
ALT	1.00 (1.00–1.01)	0.003	1.01 (1.00–1.02)	0.040
GGT	1.00 (1.00–1.00)	0.321
ALP	1.01 (1.00–1.01)	0.011	1.01 (1.00–1.01)	0.025
ALB	0.97 (0.94–1.00)	0.059	0.97 (0.93–1.00)	0.083
TBIL	1.00 (0.98–1.02)	0.717
DBIL	1.03 (0.99–1.08)	0.141
AFP	1.98 (1.40–2.80)	< 0.001	2.01 (1.41–2.87)	< 0.001

Note: Data in parentheses are 95% CIs. AST, aspartate aminotransferase; ALT, alanine aminotransferase; GGT, gamma‐glutamyl transferase; ALP, alkaline phosphatase; ALB, albumin; TBIL, total bilirubin; DBIL, direct bilirubin; AFP, alpha‐fetoprotein. *p < 0.05 was considered statistically significant.

3.5. Construction of HCC Prognostic Models

Using Cox regression, DT, RSF, and XGBoost survival algorithms, eight prognostic models were developed based on clinical and combined variables. To ensure generalizability and robust performance across different cohorts, the RSF‐based combined model was selected for final prognostic analysis (Doc S9).

3.6. Incremental Predictive Value of the 2PI System

The 2PI system provided incremental predictive value to the clinical model. DeLong's test demonstrated that, relative to the clinical‐only model, the combined model achieved generally higher C‐indices in internal and external test sets (training set C‐index, 0.76 [95% CI, 0.72–0.80] vs. 0.73 [95% CI, 0.69–0.77]; p = 0.060; internal test set C‐index, 0.69 [95% CI, 0.63–0.75] vs. 0.59 [95% CI, 0.52–0.65]; p < 0.010; external test set C‐index, 0.68 [95% CI, 0.57–0.79] vs. 0.53 [95% CI, 0.40–0.66]; p < 0.010) (Tables 4 and S8).

TABLE 4.

AUC and C‐index of the RSF combined model across all sets.

Data set	AUC(%)
	1 year	2 year	3 year	C‐index
Training	78.70 (72.78–84.62)	77.67 (71.75–83.60)	81.70 (75.67–87.74)	0.76 (0.72–0.80)
Internal test	68.68 (57.90–79.46)	72.21 (61.79–82.64)	78.88 (67.55–89.99)	0.69 (0.63–0.75)
External test	73.29 (56.74–89.84)	65.06 (55.29–89.60)	76.97 (67.91–99.99)	0.68 (0.57–0.79)

Note: Data in parentheses are 95% CIs. AUC, area under the curves; C‐index, concordance index; RSF, random survival forest.

3.7. Comparison With Radiomics‐Based Models

To further evaluate model performance, radiomics‐based comparisons were conducted, as detailed in Doc S10. The radiomics‐clinical combined Cox model demonstrated acceptable discrimination in the training set (C‐index: 0.708). However, it exhibited significant performance degradation in the independent test set, with 1‐, 2‐, and 3‐year area under the curves (AUCs) of 0.67, 0.51, and 0.55, respectively. Furthermore, its Kaplan–Meier analysis failed to effectively stratify risk groups (p > 0.05) (Figures S5 and S6). In contrast, the proposed combined model incorporating 2PI maintained higher discrimination in the external test set (1‐, 2‐, and 3‐year AUCs of 0.73, 0.65, and 0.77, respectively), and its Kaplan–Meier analysis demonstrated significant risk stratification (p < 0.05).

3.8. Model Performance and Interpretability Analyses

Figure 4 illustrates the time‐dependent ROC curves of the various ML models for predicting HCC recurrence, with the RSF‐based survival model demonstrating favorable time‐dependent AUCs across all datasets. Decision curve analysis (DCA) further confirmed the superior net clinical benefit of the combined model across a reasonable threshold probability range (Figure S7). In addition, model interpretability analyses were performed. SHAP analysis highlighted the substantial contribution of the 2PI system to model predictions. Figure 5 ranks recurrence‐related features by SHAP values, while the representative case comparisons in Figure 6 visually underscore the clinical utility of the model.

FIGURE 4.

Comparison of AUCs among the four prognostic models. Time‐dependent ROC curves for predicting RFS at 1, 2, and 3 years using Cox regression, DT, RSF, and XGBoost in the (A) training set, (B) internal test set, and (C) external test set. AUCs, area under the curves; ROC, receiver operating characteristic; RFS, recurrence‐free survival; DT, decision tree; RSF, random survival forest; XGBoost, extreme gradient boosting.

FIGURE 5.

Interpreting the RSF model with SHAP. (A) The importance of each feature ranked by mean absolute SHAP values; (B) the directional effect of predictors on the model output. In the beeswarm plot, colors represent feature values (brighter: higher values; darker: lower values), and the horizontal axis indicates the SHAP value, where a greater distance from zero reflects a larger influence on the model output.

FIGURE 6.

Illustrative cases of HCC recurrence prediction by the 2PI system. Patient A, with a 2PI score of 25, had a high risk of HCC recurrence, and the actual recurrence occurred at 30 months after surgery. Patient B, with a 2PI score of 7, had a low risk of HCC recurrence, and no recurrence was observed at 92 months after surgery.

4. Discussion

The 2PI system integrates imaging, pathological, and prognostic information with differentiated weights, enabling effective stratification of patients into high‐risk and low‐risk groups for HCC recurrence. By combining the 2PI system with conventional clinical parameters, the present study developed a composite model whose predictive performance surpassed that of traditional models relying solely on clinical data. Notably, this model demonstrated preliminary predictive capability across both RFA and SR cohorts.

Tumors consist not only of malignant cells but also of surrounding stroma and stromal cells, together forming a complex microenvironment [17]. Medical images non‐invasively capture this heterogeneity, encoding information far beyond simple diagnostic categorization. Thus, their core value must extend beyond diagnosis to inform therapeutic decisions [18]. While imaging‐derived scoring models for liver diseases are well established [16, 19], the relative importance of individual imaging features remains undefined. Determining which features exert the greatest influence, and whether they should be differentially weighted is essential for optimizing the Liver Imaging Reporting and Data System (LI‐RADS) [20]. Tumor progression is influenced by tumor subtype (e.g., MTM), MVI and the underlying liver condition with PHT [21, 22]. PHT is an independent HCC recurrence risk factor, with prognostic significance comparable to MVI, MTM, and cirrhosis [10, 23]. Accordingly, we incorporated intratumoral features, peritumoral findings, and portal venous morphology to further elucidate underlying pathophysiological mechanisms.

Dynamic CECT fulfills the standardized LI‐RADS diagnostic requirements. Given that the LR‐M category indicates significantly poorer RFS [24, 25], we re‐assessed the LR‐M category for all HCC patients; the resulting 37% incidence aligns with prior reports [26]. Nontarget features (such as infiltrative appearance and necrosis) were significantly associated with recurrence, reflecting aggressive tumor biology [27, 28]. These aggressive phenotypes are an established imaging signature of the MTM subtype [9, 29]. Among these, substantial necrosis is an independent MTM‐HCC predictor, with a specificity of 93% [11]. Consistent with this pathophysiological link, our results confirm a strong association between nonsmooth margins, LR‐M category, and MTM‐HCC.

MVI is highly prevalent in multinodular disease; thus, predicting its presence in solitary tumors carries greater clinical significance by identifying occult invasion [30]. Our findings align with previous reports [31, 32]: intratumoral arteries, nonsmooth margins, incomplete “capsule”, and corona enhancement were all linked to MVI. Importantly, corona enhancement reflects an abnormal hemodynamic microenvironment characterized by increased neovascularization or early venous drainage [30, 33], which explains its high prognostic weight (assigned 5 points) in our model. Unlike early arterial rim APHE, corona enhancement appears in the late arterial and early portal venous phases, serving as a potent indicator of poor prognosis [26]. Additionally, despite LI‐RADS morphological types predicting poorer prognosis [27, 34], morphological type was not retained in our model; however, infiltrative morphology remains a key determinant of adverse outcomes.

Underlying cirrhosis and PHT are present in most HCC patients and serve as independent risk factors for recurrence, comparable to MVI [35, 36, 37]. Although not an absolute contraindication to resection, the presence of PHT identifies a complex patient subpopulation [38]. Notably, both conditions are well‐established independent predictors of HCC recurrence [23, 39]. As a hemodynamic adaptation to elevated portal venous pressure, the development of spontaneous portosystemic shunts (SPSS) is a hallmark of decompensated cirrhosis [40]. In alignment with this, our study demonstrated that portal vein diameter and collateral circulation were significant variables, both exhibiting a strong correlation with pathological nodular cirrhosis.

In clinical practice, the 1‐, 2‐, and 3‐year time points are of great significance for HCC follow‐up. Specifically, 1 year marks the peak of early recurrence, 2 years serves as the dividing line between early and late recurrence, and a recurrence rate exceeding 50% at 3 years represents a watershed for prognostic assessment and treatment adjustment [41]. In our analysis, the 3‐year AUCs (training set AUC: 0.82; internal test set AUC: 0.79; external test set AUC: 0.77) surpassed those of the 1‐year prediction (training set AUC: 0.79; internal test set AUC: 0.69; external test set AUC: 0.73).

First, the relatively lower 1‐year AUC primarily stems from the distinct characteristics of our cohort: all patients had solitary tumors ≤ 5 cm at an early stage [42, 43]. This inherently excludes multifocal lesions, which carry a high risk of early recurrence. In such relatively early‐stage patients, early recurrence is predominantly driven by occult micrometastases. Capturing these microscopic and random events via preoperative CT is exceedingly difficult, representing an inherent physical limit for any model based on conventional macroscopic imaging. Nevertheless, the fact that the 1‐year AUC reached its current level demonstrates that the tumor‐related imaging features within our model have captured the maximal possible macroscopic risk warnings for early recurrence. Second, for patients with early‐stage HCC superimposed on a cirrhotic background, the paramount clinical challenge lies in the choice of initial treatment strategy (surgical resection vs. ablation). Long‐term (2‐ to 3‐year) recurrence is primarily driven by multicentric occurrence within the cirrhotic liver. By integrating imaging and clinical features reflective of the hepatic background, our model accurately identifies this sustained, deteriorating long‐term risk. If the model indicates a highly elevated 3‐year recurrence risk, it implies that the patient is highly susceptible to de novo tumorigenesis in the compromised liver.

A central contribution of this study is the innovative development of the 2PI system. While complex models often lack interpretability, the 2PI system demonstrates favorable simplicity and reproducibility. Importantly, the 2PI system directly and effectively stratifies patients into high‐ and low‐recurrence risk groups, revealing significant differences in gene expression profiles and immune microenvironment features between the two groups. This provides crucial insights into understanding the underlying tumor biological behavior. Furthermore, exploratory subgroup analyses suggested that the 2PI system may have some value for risk stratification across different treatment subgroups (SR and RFA); however, given that the differences between subgroups did not reach statistical significance, these findings should be interpreted only as preliminary clues and should not be regarded as evidence supporting the preferential selection of SR or RFA. Well‐designed studies with larger sample sizes, particularly prospective validation, are needed to further evaluate whether the 2PI system holds value as an aid to therapeutic decision‐making.

Another core strength of this study lies in the construction of a comprehensive prognostic prediction model incorporating the 2PI and key clinical indicators (AFP, ALT, ALP, and tumor size), which precisely serves a highly clinically representative patient population—those with early‐stage tumors coexisting with ACLD. In clinical practice, such patients face complex therapeutic decisions. Therefore, the contribution of this study lies not in pursuing a universally high C‐index, but in providing a novel prognostic prediction approach for this specific population. The results demonstrated that the comprehensive prognostic model exhibited stable performance across different test cohorts (3‐year AUC of 0.79 for the RFA cohort and 0.77 for the SR cohort). Although its statistical discrimination is moderate, DCA results revealed a positive net benefit across a broad spectrum of threshold probabilities, thereby effectively identifying patients who would benefit from interventions.

To ensure comparability between treatment groups, strict inclusion criteria were applied to focus only on solitary HCC lesions measuring ≤ 5 cm. Including larger tumors would preclude RFA (a relative contraindication for tumors > 5 cm in clinical practice), undermining our comparative aim; moreover, early‐stage HCC is where SR versus RFA decisions are most challenging and clinically impactful. Nevertheless, several limitations warrant discussion. First, a notable limitation of our study is the moderate C‐index (0.68–0.76), which restricts the model's capacity to precisely rank individual recurrence risk. This moderate discrimination likely reflects an inevitable trade‐off for model simplification, compounded by inherent retrospective bias. Second, the external test set was relatively small (n = 47), especially after excluding cases with poor image quality or tumor size > 5 cm. The cohort was collected from two sub‐centers with strict quality control to ensure consistency. While this reduced the eligible sample size, validation across distinct centers supports preliminary generalizability. Future prospective validation in larger, multicenter cohorts is needed, and we are currently collecting data from additional centers in an ongoing study to further validate the model. Third, patients treated with RFA lacked pathological confirmation. Although a strict clinical diagnosis was applied and further validated by follow‐up outcomes, the absence of a pathological gold standard still implies a potential risk of minor misclassification. Future research should aim to incorporate prospective cohorts with pathological confirmation to further validate and optimize the model. Fourth, this study relied solely on a CT‐based model; future cross‐validation using MRI could further corroborate these findings. Fifth, tumor location was not incorporated into the analysis; further studies will evaluate tumor‐vessel proximity and other spatial relationships. Sixth, the relatively short follow‐up duration limits the evaluation of long‐term prognosis beyond 3 years. Due to enrollment concluding in 2022, 5‐year data are currently unavailable, precluding the observation of very late recurrence patterns. Future studies with extended follow‐up are warranted to validate long‐term efficacy.

In conclusion, the 2PI‐based RSF model demonstrates strong usability and interpretability (Figure 7). Integrated into routine clinical diagnostics, it may serve as a promising non‐invasive imaging biomarker for HCC risk stratification and a practical preoperative prognostic tool.

FIGURE 7.

Graphical abstract in this study. Using differential weighting, we developed a 2PI‐based risk stratification system for HCC recurrence. A comprehensive prognostic model incorporating clinical variables and the 2PI score further enhanced postoperative prediction in solitary HCC and showed initial generalizability across SR and RFA cohorts.

Author Contributions

Yingzhu Cui: writing – original draft, writing – review and editing, conceptualization, methodology, investigation, formal analysis. Dezhi Zhang: writing – review and editing, methodology, investigation, formal analysis. Fan Lin: writing – review and editing, methodology, investigation, formal analysis. Liang Guo: writing – review and editing, methodology, investigation, formal analysis. Lei Zhang: writing – review and editing, methodology, investigation, formal analysis. Kan He: writing – review and editing, formal analysis, investigation, funding acquisition. Hongbin Wang: writing – review and editing, investigation, formal analysis. Shuang Zheng: writing – review and editing. Kun Zhang: writing – review and editing. Yu Fu: writing – review and editing, conceptualization, supervision. Huimao Zhang: supervision, resources, funding acquisition, investigation, formal analysis, methodology, conceptualization, writing – review and editing.

Funding

This study was funded by the National Natural Science Foundation of China (Grant Nos. 82202256, 82430062, 12426313, U22A20351), the Jilin Provincial Healthcare Talent Special Program (JLSWSRCZX2023‐3), and the National Key Research and Development Program of China (No. 2024YFE0213800).

Ethics Statement

This multicenter retrospective study was approved by the ethics committees of three tertiary medical centers (The First Hospital of Jilin University, Tianjin First Central Hospital, and Yantai Yuhuangding Hospital), with a waiver of informed consent (AAF‐IRB‐029‐06).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: cas70519‐sup‐0001‐Supinfo.docx.

Doc S1. CT scanning examination.

Doc S2. Definition of imaging features.

Doc S3. Image quality assessment and consistency test.

Doc S4. One‐to‐one validation between imaging and pathological findings.

Doc S5. Follow‐up protocol.

Doc S6. Combining exp(β) and Kendall's tau‐b coefficient: theoretical considerations.

Doc S7. Sample size calculation.

Doc S8. 2PI risk stratification‐associated gene expression profiles and immune microenvironment features.

Doc S9. Model selection.

Doc S10. Radiomics model development process.

Table S1: CT scanning and imaging acquisition protocol.

Table S2: Calculation and comparison of contrast‐to‐noise ratio.

Table S3: Characteristics of patients grouped based on pathologic features.

Table S4: Kendall's tau‐b correlation coefficient.

Table S5: Image scoring based on the 2PI scoring system.

Table S6: Sensitivity analysis of weighting strategies on model performance.

Table S7: Multivariable Cox proportional hazards regression.

Table S8: C‐index performance of two models across three cohorts.

Figure S1: Comparison of intra‐ and inter‐group kappa values for imaging features.

Figure S2: Impact of comorbidities on the recurrence of HCC.

Figure S3: Biological basis underlying 2PI risk stratification.

Figure S4: Exploratory subgroup analysis of RFS stratified by the 2PI system.

Figure S5: LASSO regression and bootstrap stability analysis of the model C‐index.

Figure S6: Kaplan–Meier analysis of recurrence‐free survival and ROC curves for the radiomics‐clinical combined model in the training and test sets.

Figure S7: Multi‐time‐point decision curve analysis of the random survival forest model.

Data Availability Statement

The data and materials used in this study are available upon reasonable request from the corresponding author, in accordance with the privacy policy and ethical approval.