HepaKineticNet: Deep Learning of CT-Derived Intratumoral Heterogeneity for Overall Survival Prediction in BCLC Stage B-C Hepatocellular Carcinoma Treated with Locoregional Therapy Plus First-Line TKI and Anti–PD-1 Therapy

Abstract

Purpose

To develop HepaKineticNet, a deep learning framework that models inter-phase kinetic dynamics on multi-phase contrast-enhanced CT (CECT) to generate intratumoral heterogeneity (ITH) maps, and to assess in an independent external cohort whether a clinical–radiomics fusion model improves overall survival (OS) prediction in BCLC stage B–C hepatocellular carcinoma (HCC) treated with locoregional therapy plus first-line TKI and anti–PD-1 agents.

Patients and Methods

In this bi-institutional retrospective study, HepaKineticNet was developed using 359 patients from one institution (231 training, 128 validation) and evaluated in an independent external test cohort of 95 patients from a second institution. HepaKineticNet integrates a Spatial Kinetic Differential Attention module into an nnU-Net backbone, separately encoding arterial wash-in and delayed wash-out kinetics, with multiscale coefficient-of-variation self-supervision. Ten radiomics features (hierarchical filtering) and three clinical variables selected by LASSO-penalised Cox regression (10-fold cross-validation) were used to build clinical, radiomics, and fusion Cox models.

Results

A total of 277 deaths occurred, 128 of them in the development cohort. Baseline characteristics were balanced (all P>0.05). In the external test cohort, the fusion model achieved a C-index of 0.713, exceeding the clinical model (0.651; P=0.012). Time-dependent AUCs at 1, 2, and 3 years were 0.760 (95%CI 0.651–0.864), 0.838 (95%CI 0.736–0.930), and 0.909 (95%CI 0.686–1.000), whereas those of the clinical model were 0.676, 0.742, and 0.772. Kaplan–Meier analysis showed median OS of 9.5 vs 19.0 months for high- vs low-risk strata (HR 3.02, 95%CI 1.86–4.91; P<0.001). Discrimination was consistent across ten prespecified subgroups.

Conclusion

HepaKineticNet-derived ITH maps encode kinetic and spatial information that modestly improves individualized OS prediction and risk stratification in this combination-therapy HCC population, and may provide a noninvasive decision-support adjunct pending prospective validation.

Introduction

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality globally and is endemic in hepatitis B virus (HBV)-prevalent regions of East Asia.¹ Despite expanded therapeutic options-from surgical resection and locoregional therapies to immune checkpoint inhibitors (ICIs) and tyrosine kinase inhibitors (TKIs)-accurate individualized prognostication remains critical for treatment selection, patient counseling, and clinical trial design. Current staging algorithms, including the Barcelona Clinic Liver Cancer (BCLC) system, Child-Pugh score, and albumin–bilirubin (ALBI) grade, are validated for population-level risk stratification but are insensitive to the intratumoral spatial heterogeneity (ITH) that governs invasive behavior, immune escape, and systemic resistance.^2,3

Multi-phase CECT is the global imaging standard for HCC staging.⁴ The arterial wash-in and delayed-phase wash-out directly reflect microvascular density and contrast agent clearance-noninvasive surrogates for the hemodynamic microenvironment driving ITH.⁵ Yet conventional radiomics pipelines extract features from static or naively concatenated multi-phase images, discarding the inter-phase temporal dynamics most informative for perfusion heterogeneity characterization.⁶ Existing deep learning segmentation networks typically treat multi-phase stacks as a static multi-channel input, ignoring the directional kinetic change between phases.^7,8 Published HCC radiomics OS models report C-indices predominantly in the 0.70–0.80 range,^9,10 and to our knowledge, no framework explicitly integrates hemodynamic priors into heterogeneity estimation.

Radiomic and deep-learning models for overall survival in HCC report C-indices of approximately 0.70–0.80, but almost exclusively in single-centre cohorts of resected, early-stage disease; performance in intermediate-to-advanced disease treated with combined locoregional and systemic therapy remains poorly characterised. Previous approaches to spatial heterogeneity have relied largely on habitat imaging, in which the tumour is clustered into a small number of discrete subregions, whereas kinetic behaviour varies continuously across the lesion. Both conventional radiomics and habitat analysis, moreover, characterise the tumour at a single pre-treatment time point — a constraint shared by the present work and one that longitudinal delta-radiomics is intended to address.

We introduce HepaKineticNet, an end-to-end deep learning framework that addresses this gap through three innovations: (1) a Spatial Kinetic Differential Attention (SpatialKDA) module that constructs explicit phase-difference representations capturing arterial wash-in and delayed wash-out independently; (2) a multiscale coefficient-of-variation (CV) self-supervision strategy enabling physics-constrained ITH mapping without voxel-level annotations; and (3) a clinical–radiomics fusion Cox model validated in an independent bi-institutional cohort. Our objectives were to develop a kinetics-aware segmentation-heterogeneity joint network, derive a validated radiomics panel from ITH maps, and evaluate whether the fusion model improves individualized HCC OS prediction over conventional clinical staging.

Materials and Methods

Study Design and Patients

A total of 521 patients with newly diagnosed hepatocellular carcinoma who received interventional locoregional therapy combined with first-line targeted therapy and immune checkpoint inhibitors were initially screened at two independent tertiary medical centers between January 2019 and December 2023. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study protocol was reviewed and approved by the Ethics Committee of Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences (approval no. SDTHEC202408020) and by the Ethics Committee of Yantai Yuhuangding Hospital, Qingdao University (approval no. K2024-044). Given the retrospective design of the study, the requirement for written informed consent was waived by both ethics committees. Inclusion criteria were: (1) confirmed HCC by histopathology or meeting the 2018 AASLD/EASL clinical imaging diagnostic criteria;⁴ (2) complete triphasic contrast-enhanced CT (arterial, portal-venous, and delayed phases) performed within 30 days prior to the initiation of any anti-tumor treatment; (3) availability of complete baseline clinical data and follow-up information with a minimum follow-up duration of one month; and (4) at least one measurable intrahepatic target lesion assessable per RECIST version 1.1. Exclusion criteria were: (1) suboptimal image quality precluding reliable tumor delineation and feature extraction; (2) any prior anti-tumor therapy before the index contrast-enhanced CT examination; (3) concurrent secondary malignancy; and (4) critical missing data in key clinical or imaging variables. Based on these eligibility criteria, 454 consecutive patients were ultimately enrolled. The study flowchart is shown in Figure 1.

Figure 1.

Flowchart of patient screening, eligibility assessment, and final enrollment, showing development of the model in the Shandong Cancer Hospital cohort (training n=231, validation n=128) and independent external testing in the Yantai Yuhuangding Hospital cohort (n=95).

To ensure data consistency across institutions, a unified data collection protocol was implemented at both centers, with standardized definitions of all clinical variables and follow-up endpoints established a priori. All clinical and imaging data were independently verified by two dedicated researchers, and any discrepancies were resolved by consensus.

The prognostic model was developed exclusively in the Shandong Cancer Hospital cohort, which was divided into a training set (n=231) and an internal validation set (n=128) by stratified random allocation on BCLC stage and ALBI grade. All 95 patients from Yantai Yuhuangding Hospital were reserved as an independent external test set: no data or outcome information from this centre contributed to feature selection, estimation of normalisation parameters, LASSO or Cox model fitting, or definition of the risk threshold. The upstream segmentation and heterogeneity-mapping network was trained by five-fold cross-validation on pooled images from both centres; this stage used only tumour masks and self-supervised targets derived from image intensities, and no survival or outcome data of any kind. The study is reported in accordance with the TRIPOD+AI and CLAIM recommendations (Online Resource Checklist S1).

Treatment Protocol

All transarterial chemoembolization (TACE) procedures were performed via percutaneous femoral artery access using the Seldinger technique. Under digital subtraction angiography (DSA) fluoroscopic guidance, selective catheterization was sequentially carried out in the celiac trunk, proper hepatic artery, and superior mesenteric artery to characterize tumor-feeding vessels. When hepatic arteriovenous fistulae were identified on pre-procedural superselective angiography, the following strategy was applied: if the microcatheter tip could be advanced distal to the fistula opening, embolization of tumor-feeding arteries was prioritized before the fistula was sealed; if the microcatheter tip could not traverse the fistula, the fistula was occluded first, followed by complete embolization of tumor-feeding vessels. All procedures in this study employed conventional TACE (cTACE) methodology. The lipiodol emulsion was prepared by thoroughly mixing epirubicin (40 mg), cisplatin (60 mg), and lipiodol (5–10 mL), with the final lipiodol volume individualized according to tumor size and vascularity assessed on angiography. Fistula occlusion was achieved using calibrated microspheres, gelatin sponge particles, microcoils, and/or polyvinyl alcohol particles, selected individually or in combination based on hemodynamic parameters. These agents were used solely to occlude hepatic arteriovenous fistulae and served neither as drug carriers nor as the embolic agent for tumor treatment; drug-eluting bead TACE was not performed in any patient. TACE cycles were repeated at 3–4-week intervals, with the number of cycles determined by clinical response and hepatic functional reserve.

In the combination therapy group, hepatic arterial infusion chemotherapy (HAIC) was administered concurrently using a FOLFOX-based regimen delivered via intra-arterial infusion through a microcatheter secured within the target hepatic artery: oxaliplatin (85 mg/m² over 2 hours), leucovorin (400 mg/m² over 2 hours), followed by a 5-fluorouracil (5-FU) bolus (400 mg/m² over 1 hour) and continuous 5-FU infusion (2400 mg/m² over 23 hours). Combined TACE-HAIC sessions were repeated at 3–4-week intervals. Chemotherapeutic dosages were adjusted as appropriate based on the patient’s general condition, body surface area, and angiographic findings.

Programmed cell death protein 1 (PD-1) inhibitors (200 mg per agent) were administered intravenously beginning on post-procedure day 1 at 3–4-week intervals; available agents included sintilimab, tislelizumab, and camrelizumab. Tyrosine kinase inhibitors (TKIs) were initiated from post-procedure day 3 and included sorafenib (400 mg twice daily), donafenib (400 mg twice daily), or lenvatinib (12 mg/day for body weight ≥ 60 kg; 8 mg/day for < 60 kg). For grade 3–4 treatment-related adverse events, sorafenib and donafenib were reduced to once daily, while lenvatinib was adjusted to 8 mg/day (≥ 60 kg) or 4 mg/day (< 60 kg). Because no evidence supports the superiority of any individual PD-1 inhibitor or TKI, agent selection was primarily determined by patient preference, drug acquisition costs, and local health insurance reimbursement policies.

Image Preprocessing and Segmentation

All CECT volumes underwent a standardized preprocessing pipeline: (1) resampling to isotropic 1×1×1 mm³ voxels; (2) intensity normalization (1st–99^th percentile windowing); (3) deformable registration of the portal-venous and delayed phases to the arterial reference frame using the DEEDS algorithm (dense displacement sampling with MIND self-similarity descriptors); and (4) assembly into a three-channel volumetric tensor. Full registration parameters are given in Online Resource Supplementary Methods S1. Tumor segmentation ground truth was generated independently by two radiologists (>5 and >8 years hepatic imaging experience), with discordant cases resolved by consensus or a third expert reader. Inter-observer agreement between the two independent readers was substantial (Dice 0.879). Automated segmentation by the HepaKineticNet backbone achieved a Dice coefficient of 0.864 against the consensus reference.

HepaKineticNet Architecture

HepaKineticNet follows the self-configuring nnU-Net encoder–decoder backbone,¹¹ which progressively compresses spatial information through residual convolutional blocks and recovers resolution via skip-connected upsampling. A dedicated heterogeneity estimation branch at the network bottleneck generates a continuous spatial ITH map in parallel with the segmentation mask. Because this estimator operates at an effective resolution on the order of a few voxels and its output is upsampled to the image grid, the ITH map represents a regional rather than a voxel-level characterisation of kinetic dispersion. The SpatialKDA module separates three-phase feature maps and constructs two kinetic difference representations-F_AP→PVP (arterial wash-in) and F_PVP→DP (delayed wash-out)-processed by a lightweight spatial attention sub-network¹² to generate phase-transition attention weights that emphasize hemodynamically active regions. Attention-weighted kinetic features are fused with static encoder representations via element-wise addition. Multiscale CV targets computed at 3×3×3, 5×5×5, and 9×9×9 voxel kernels provide self-supervised regression objectives. Multi-resolution auxiliary branches at three decoder stages, together with an image-space consistency loss penalizing deviation from raw inter-phase CT signal differences and an uncertainty-error consistency loss, constrain training and ground the learned heterogeneity in measurable imaging signal. Complete loss definitions and weights are given in Online Resource Supplementary Methods S1, and the full training configuration is summarised in Online Resource Table S4. The overall architecture of HepaKineticNet and the analytical workflow are illustrated in Figure 2.

Figure 2.

Architecture of HepaKineticNet and overall analytical workflow. (A) Multi-phase CECT input: arterial, portal-venous and delayed phase images acquired in the routine imaging workflow, with the tumour region of interest outlined. (B) Kinetic differential modelling: the SpatialKDA module encodes directional wash-in (arterial→delayed) and wash-out enhancement changes through phase-difference attention, emphasising haemodynamically active tumour regions. (C) Network structure: an nnU-Net encoder–decoder backbone with the SpatialKDA module at the bottleneck, trained with multiscale coefficient-of-variation self-supervision together with image-space and uncertainty consistency losses, producing a tumour segmentation mask and a continuous ITH map in parallel. (D) Schematic illustration of the continuous ITH map, indicating hotspots, their spatial distribution, the core-to-rim gradient and boundary proximity; actual HepaKineticNet ITH maps from two representative patients are shown in Figure 3. (E) Feature and fusion modelling: ten ITH-derived radiomics features and three LASSO-selected clinical variables are combined in a fusion Cox model. (F) Individualized overall survival prediction: the resulting risk score stratifies patients into low- and high-risk groups (external test cohort: C-index 0.713; 1-, 2- and 3-year AUC 0.760, 0.838 and 0.909; median overall survival 19.0 versus 9.5 months).

Abbreviations: CECT, contrast-enhanced computed tomography; ITH, intratumoral heterogeneity; SpatialKDA, spatial kinetic differential attention.

Feature Extraction and Selection

From HepaKineticNet ITH maps, 172 candidate radiomics features were extracted across four families of operators. These were not obtained with PyRadiomics; they were computed with an in-house Python pipeline (Python 3.10; SimpleITK 2.3, NumPy 1.26, SciPy 1.11, scikit-image 0.22) using operators designed for the spatial organisation of a continuous heterogeneity field, for which there are no direct equivalents in the IBSI standard feature set. Implementation details are given in Online Resource Supplementary Methods S1.

The ITH map is a continuous, normalised field in which higher values denote greater local kinetic non-uniformity (Figure 3). Habitat analysis in the conventional sense -unsupervised clustering of the tumour into discrete subregions- was not performed; hotspots were defined operationally by the P90 and P95 percentile thresholds of this continuous field and are not intended to represent discrete biological habitats.

Figure 3.

Representative HepaKineticNet ITH maps in two patients with contrasting outcomes. Upper row, a high-risk patient (risk score 8.75; overall survival 13 months, died); lower row, a low-risk patient (risk score 0.48; alive at 30 months). From left to right: arterial, portal-venous and delayed phase images with the tumour outline; AP→PVP (wash-in) and PVP→DP (wash-out) phase-difference maps on a shared HU scale; and the continuous HepaKineticNet ITH map. ITH maps are displayed on a shared colour scale (gamma = 0.5); solid and dashed contours mark the P90 and P95 hotspot thresholds, and the corresponding within-tumour percentile values are annotated in each panel. Higher ITH values denote greater local kinetic non-uniformity. The ITH map is generated at bottleneck resolution and upsampled, and so depicts regional aggregation of the voxel-level dynamics shown in the difference panels; it is a continuous field rather than a partition into discrete habitats.

Hotspot fractional volumes. These features quantify the extent of extreme kinetic dispersion within the tumour: the proportion of tumour voxels whose ITH values exceed the P90 or P95 percentile of the lesion’s own distribution, together with the mean volume of the connected components forming the P95 hotspot region. Larger values indicate that a greater share of the tumour behaves in a kinetically divergent manner, which may reflect regions of markedly different perfusion.

Spatial organisation metrics. These describe where the hotspot micro-domains lie rather than how large they are, and comprise the mean and standard deviation of the distances between the centroids of hotspot connected components, the mean shortest distance from those components to the tumour boundary, and the displacement of the hotspot centroid from the tumour centroid normalised by tumour radius. Together they distinguish tumours in which kinetically divergent domains are dispersed throughout the lesion from those in which such domains are clustered towards the periphery. In the present cohort, greater inter-centroid dispersion (HR 1.34 per 1 SD) and shorter distances to the tumour boundary (HR 0.62 per 1 SD increase in distance) were both associated with shorter survival, a pattern consistent with -though not evidence for -the hypothesis that activity at the invasive front carries prognostic weight.

Radial gradient features. The ITH values sampled along rays running from the tumour centroid to its surface were fitted by linear regression, and the slope of that fit was retained, summarising the core-to-rim transition of kinetic heterogeneity in a single value.

Global statistical descriptors. These characterise the ITH field of the whole tumour without reference to location, and comprise the coefficient of variation of ITH values across the lesion and a measure of local intensity contrast within the field. Feature selection followed a three-stage pipeline: (1) Pearson correlation filtering (|r|>0.85; 172→64 features); (2) univariable Cox regression (P<0.10; 64→28); and (3) Cox-LASSO with 10-fold cross-validation (28→10 retained). The HepaKineticNet radiomics risk score was computed as the linear predictor of the LASSO-Cox model: RiskScore = Σ β_i × z_i, where z_i is the feature standardised to the training-cohort mean and standard deviation and β_i its LASSO coefficient; all ten coefficients are given in Online Resource Table S1. The fusion model risk score was obtained analogously from the Cox model containing the radiomics risk score and the selected clinical variables.

Clinical variables were selected using LASSO-penalised Cox regression with 10-fold cross-validation from a prespecified set of 15 established prognostic factors; three variables were retained: extrahepatic metastasis, CRAFITY score, and alpha-fetoprotein (AFP). The radiomics feature pipeline used the same LASSO-penalised Cox selection strategy, so that both feature sets were derived with an identical, reproducible selection framework.

Statistical Analysis

Three Cox proportional-hazards models were constructed: a clinical model (three variables selected by LASSO-Cox), a HepaKineticNet radiomics model (ten features), and a fusion model combining the radiomics risk score with the three clinical variables. Proportional hazards assumptions were verified by Schoenfeld residuals. Discriminative performance was assessed by concordance index (C-index) and time-dependent AUC at 1, 2, and 3 years using the inverse probability of censoring weighting (IPCW) estimator. Calibration was evaluated by calibration curves and Brier scores. Clinical utility was assessed by decision curve analysis (DCA). Overall survival was defined as the interval from the start of treatment to death from any cause. Because every death counts as an event under this definition, no competing event arises for the primary endpoint, and subdistribution hazard models were therefore not applicable. Kaplan–Meier analysis with log-rank test evaluated risk stratification. The risk threshold was prespecified as the median risk score of the training cohort and applied unchanged to the internal validation and external test cohorts; it was not re-optimised on test data. Dichotomisation was used only to display risk stratification, and the primary performance measures (C-index and time-dependent AUC) are threshold-free. All metrics used 1000-iteration bootstrap for 95% confidence intervals; model comparisons used the DeLong test for time-dependent AUC, and a bootstrap test for differences in C-index, with Benjamini-Hochberg FDR correction. Baseline characteristics, categorized at predefined clinical thresholds, were compared across the three cohorts using the chi-squared or Fisher’s exact test. P<0.05 was considered statistically significant. A total of 277 deaths occurred in the full cohort, of which 128 were observed in the training cohort. The fusion model comprised four predictors, corresponding to 32 events per variable during model development, above the conventional minimum of ten events per variable. No outcome information was used at any stage of network training: supervision consisted only of tumour masks and self-supervised targets derived from image intensities. Analyses were performed in R (v4.3.1; survival, timeROC, dcurves) and Python (v3.10; scikit-survival).

Results

Patient Characteristics

Of 454 enrolled patients, 232 (51.1%) were male and 360 (79.3%) were aged 50 years or older. Baseline characteristics were well balanced across the three cohorts (all P>0.05; Table 1). Overall, 373 patients (82.2%) were HBV-positive and 249 (54.8%) had underlying cirrhosis, consistent with the endemic HBV-related etiology; 373 (82.2%) were Child–Pugh class A and 369 (81.3%) had an ECOG performance status of 0. By BCLC stage, 183 patients (40.3%) had stage B and 271 (59.7%) stage C disease. A maximal tumor diameter below 5 cm was present in 294 patients (64.8%), and 280 (61.7%) had a solitary lesion. Vascular invasion was identified in 195 patients (43.0%) and extrahepatic metastasis in 164 (36.1%), spanning intermediate-to-advanced stages with marked biological heterogeneity.

Table 1.

Baseline Characteristics of Patients in the Training, Validation, and Test Cohorts

Characteristic	Training Cohort (n = 231)	Validation Cohort (n = 128)	Test Cohort (n = 95)	χ2	P value
Age (years)				1.770	0.413
<50	51 (22.1)	28 (21.9)	15 (15.8)
≥50	180 (77.9)	100 (78.1)	80 (84.2)
Sex				3.905	0.142
Female	108 (46.8)	59 (46.1)	55 (57.9)
Male	123 (53.2)	69 (53.9)	40 (42.1)
ALBI grade				3.841	0.147
0	126 (54.5)	56 (43.8)	48 (50.5)
1 and 2	105 (45.5)	72 (56.3)	47 (49.5)
AFP (ng/mL)				4.197	0.123
<400	137 (59.3)	63 (49.2)	48 (50.5)
≥400	94 (40.7)	65 (50.8)	47 (49.5)
CRAFITY score				5.210	0.074
0	144 (62.3)	89 (69.5)	71 (74.7)
1 and 2	87 (37.7)	39 (30.5)	24 (25.3)
Child–Pugh class				0.917	0.632
A	186 (80.5)	108 (84.4)	79 (83.2)
B	45 (19.5)	20 (15.6)	16 (16.8)
Cirrhosis				0.927	0.629
No	101 (43.7)	57 (44.5)	47 (49.5)
Yes	130 (56.3)	71 (55.5)	48 (50.5)
ECOG PS				3.695	0.158
0	194 (84.0)	97 (75.8)	78 (82.1)
1	37 (16.0)	31 (24.2)	17 (17.9)
Extrahepatic metastasis				5.741	0.057
No	148 (64.1)	90 (70.3)	52 (54.7)
Yes	83 (35.9)	38 (29.7)	43 (45.3)
BCLC stage				0.210	0.900
B	93 (40.3)	50 (39.1)	40 (42.1)
C	138 (59.7)	78 (60.9)	55 (57.9)
HBV infection				5.764	0.056
Negative	51 (22.1)	17 (13.3)	13 (13.7)
Positive	180 (77.9)	111 (86.7)	82 (86.3)
NLR				1.956	0.376
<3	120 (51.9)	58 (45.3)	51 (53.7)
≥3	111 (48.1)	70 (54.7)	44 (46.3)
PLR				3.224	0.200
<190	167 (72.3)	81 (63.3)	67 (70.5)
≥190	64 (27.7)	47 (36.7)	28 (29.5)
Maximum tumordiameter (cm)				2.731	0.255
<5	158 (68.4)	78 (60.9)	58 (61.1)
≥5	73 (31.6)	50 (39.1)	37 (38.9)
Tumor number				4.788	0.091
Solitary	153 (66.2)	70 (54.7)	57 (60.0)
Multiple	78 (33.8)	58 (45.3)	38 (40.0)
Vascular invasion				3.809	0.149
No	142 (61.5)	68 (53.1)	49 (51.6)
Yes	89 (38.5)	60 (46.9)	46 (48.4)

Notes: Data are presented as n (%). P values were calculated using the χ2-test.

Abbreviations: AFP, α-fetoprotein; ALBI, albumin–bilirubin; BCLC, Barcelona Clinic Liver Cancer; CRAFITY, C-reactive protein and α-fetoprotein in immunotherapy; ECOG PS, Eastern Cooperative Oncology Group performance status; HBV, hepatitis B virus; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio.

Feature Selection Results

Among 15 candidate clinical variables, extrahepatic metastasis showed the strongest univariable association with OS (HR 1.501, 95%CI 1.038–2.171; P=0.032; Online Resource Table S1). CRAFITY score (HR 1.443; P=0.058) and AFP (HR 1.281; P=0.198) showed consistent risk-elevating trends. Clinical variable selection by LASSO-penalised Cox regression with 10-fold cross-validation retained three of the 15 candidates-extrahepatic metastasis, CRAFITY score, and AFP -corresponding to the strongest univariable associations. Radiomics feature selection reduced 172 candidates to 10 retained features (Online Resource Table S1) spanning hotspot fractional volumes (hotspot_p95_ratio, hotspot_p90_ratio, hotspot_p95_mean_size), spatial organisation metrics (hotspot_distance_std, hotspot_mean_distance, hotspot_to_boundary_mean, hotspot_centroid_dist_normalized), radial gradient features (radial_gradient), and global statistical descriptors (texture_contrast, hetero_all_cv), using the same four families defined in the Methods. These features collectively capture the fractional extent of maximal-heterogeneity micro-domains, their geometric arrangement within the tumor, and radial core-to-rim architectural transitions.

Model Performance

C-indices are summarized in Table 2 and Online Resource Fig. S1. In the independent external test cohort from the second institution, the fusion model (C-index 0.713) significantly outperformed the clinical model (0.651; P=0.012). The HepaKineticNet radiomics model (0.705) exceeded the clinical model across all cohorts. The fusion model maintained stable C-indices across training (0.744), validation (0.710), and test (0.713) cohorts, indicating minimal overfitting. Time-dependent AUCs (Table 2; ROC curves in Figure 4) in the test cohort showed the fusion model exceeding the clinical model at all three landmarks (1 year: 0.760 vs 0.676; 2 years: 0.838 vs 0.742; 3 years: 0.909 vs 0.772) and the HepaKineticNet model at 2 and 3 years (0.808 and 0.831); at 1 year the HepaKineticNet model was numerically higher (0.809 vs 0.760).

Table 2.

Model Performance: C-Index and Time-Dependent AUC

Model	C-index	1-yr AUC (95%CI)	2-yr AUC (95%CI)	3-yr AUC (95%CI)
Clinical	0.675	0.658 (0.536–0.780)	0.629 (0.498–0.760)	0.612 (0.471–0.753)
HepaKineticNet	0.705	0.721 (0.608–0.834)	0.695 (0.568–0.822)	0.668 (0.528–0.808)
Fusion model (test)*	0.722	0.768 (0.671–0.866)	0.825 (0.699–0.926)	0.890 (0.680–1.000)
Fusion model (training)	0.747	0.799 (0.742–0.846)	0.817 (0.753–0.883)	0.890 (0.796–0.963)
Fusion model (validation)	0.719	0.765 (0.652–0.858)	0.823 (0.742–0.899)	0.803 (0.695–0.903)

Notes: *Significantly superior to clinical model (P=0.023, bootstrap test for the C-index difference, FDR-corrected). AUC area under time-dependent ROC curve, CI confidence interval, HR hazard ratio.

Figure 4.

Time-dependent receiver operating characteristic (ROC) curves of the fusion model for 1-, 2-, and 3-year overall survival in the training, validation, and external test cohorts; areas under the curve (AUC) with 95% confidence intervals are shown.

To characterize discrimination across the entire follow-up period rather than at fixed landmarks, time-dependent AUC (tdAUC) was computed continuously (Online Resource Fig. S2). The fusion model maintained the highest mean tdAUC in every cohort (training 0.812, validation 0.759, test 0.770), exceeding the HepaKineticNet radiomics model (0.788, 0.748, 0.766) and the clinical model (0.639, 0.601, 0.678), indicating stable long-term predictive performance.

Survival Stratification

Kaplan–Meier curves for all three models demonstrated significant separation between risk groups in the test cohort (all P<0.001; Figure 5). The fusion model achieved the sharpest stratification: median OS of 9.5 months (high risk) vs 19.0 months (low risk) and HR 3.02 (95%CI 1.86–4.91), exceeding the clinical model (HR 2.33, 95%CI 1.45–3.74) and HepaKineticNet radiomics model (HR 3.40, 95%CI 2.03–5.69). Curves showed persistent separation throughout follow-up without crossing, indicating stable prognostic discrimination over time.

Figure 5.

Kaplan–Meier overall survival curves for high- vs low-risk groups defined by the fusion-model median risk score, with numbers at risk shown beneath each panel, in the training, validation, and external test cohorts (log-rank P < 0.001).

Calibration, DCA, and Multivariable Analysis

Calibration curves confirmed close agreement between fusion model predicted and observed survival (1-year slope 0.92, 2-year 0.88, 3-year 0.85; Online Resource Fig. S3). The clinical model showed systematic overestimation of mortality risk in high-risk deciles. In the training and validation cohorts the fusion model provided higher net benefit than the clinical model across most threshold probabilities; in the external test cohort, net benefit was comparable to or slightly below that of the clinical model at thresholds above approximately 0.25 (Figure 6). In the fusion multivariable Cox model (Online Resource Table S2), the HepaKineticNet radiomics risk score remained the strongest predictor (HR 1.695, 95%CI 1.509–1.903; P<0.001), followed by extrahepatic metastasis (HR 1.696, 95%CI 1.178–2.443; P=0.005); CRAFITY score (HR 1.398, 95%CI 0.962–2.033; P=0.079) and AFP (HR 1.344, 95%CI 0.943–1.916; P=0.102) were not independently significant.

Figure 6.

Decision curve analysis for 1-, 2-, and 3-year overall survival prediction by the three models in the training, validation, and external test cohorts.

Subgroup Analysis

Fusion model C-indices across ten prespecified subgroups ranged from 0.698 (multiple tumors) to 0.752 (ALBI grade 1), exceeding the imaging-only model in every subgroup (Online Resource Table S3). Notably, the model maintained robust discrimination among lower-burden subgroups (solitary tumor ≤5 cm, C-index 0.735; ALBI grade 1, 0.752), where macroscopic staging provides the least differentiation-precisely the patient group where refined prognostic tools offer the greatest clinical value.

Discussion

This study demonstrates that HepaKineticNet-derived ITH maps, generated through explicit kinetic differential modeling of multi-phase CECT, encode prognostically meaningful spatial biological information beyond conventional macroscopic staging. The fusion model achieved a C-index of 0.713 and a hazard ratio of 3.02 between risk strata in an independent external cohort, with consistent performance across cohorts and ten prespecified subgroups. Although the gain in C-index over the clinical model was modest (Δ = 0.062; P = 0.012), the C-index summarises rank concordance over the entire follow-up and is insensitive to improvements concentrated at clinically relevant landmarks; the larger differences in 2-year (0.838 vs 0.742) and 3-year (0.909 vs 0.772) time-dependent AUC, together with the stronger separation between risk strata (HR 3.02 vs 2.33), suggest that the incremental information is clinically non-trivial.

The superiority of the radiomics model over the clinical model reflects a core limitation of macroscopic staging: it collapses diverse ITH into a single class, obscuring microstructural determinants of aggressiveness. Clinical variables such as AFP, ALBI grade, and vascular invasion reflect tumor-host interaction outcomes rather than the spatial organization of malignant micro-domains.^13,14 The dominant retained features—hotspot P95 fractional volumes and boundary-proximity metrics-may correspond to high-density neoangiogenic zones at the invasion front,^5,15 an interpretation that is consistent with published radiogenomic data but was not verified histologically here and should be considered hypothesis-generating. Radial-gradient features encoding core-to-rim transitions are consistent with ecological models in which peripheral subpopulations may drive metastatic seeding,¹⁶ and inter-hotspot dispersion metrics accord with the view that the architectural arrangement of heterogeneous subregions contributes to malignant potential beyond aggregate magnitude.^17,18

Two dimensions of the heterogeneity field should be distinguished when these features are interpreted. The magnitude of ITH values within a lesion, which is visibly higher in the high-risk than in the low-risk example of Figure 3, is not the same quantity as the fractional volume occupied by hotspots, and in our data the two carried associations of opposite sign: larger hotspot fractional volumes were associated with lower risk (hotspot_p95_ratio, HR 0.70 per 1 SD), whereas greater dispersion of hotspot components (HR 1.34 and 1.44) and their proximity to the tumour boundary (HR 0.62 per 1 SD increase in distance) were associated with higher risk. One reading is that prognosis relates less to how much of the tumour is kinetically divergent than to how that divergence is spatially arranged, fragmented and peripherally situated domains constituting the adverse pattern while a large but consolidated central hotspot does not. This interpretation is post hoc; the fractional-volume coefficients were estimated under LASSO shrinkage in the presence of correlated spatial features, and the sign of this association requires confirmation in an independent cohort before it is relied upon.

The fusion model added little to the imaging-only model in the external cohort (C-index 0.713 vs 0.705), and the imaging-only model achieved a numerically larger separation between risk strata (HR 3.40 vs 3.02). Most of the incremental information therefore resides in the imaging features themselves rather than in their combination with clinical variables, and the case for combining them rests on conceptual complementarity rather than on a demonstrated performance gain: clinical variables encode systemic host-tumor state (inflammation, tumor dissemination), whereas HepaKineticNet features capture local intra-lesional microarchitecture. Their integration reflects both systemic state and local phenotype, consistent with multimodal integration improving precision-oncology risk stratification.¹⁹ The attenuation of the CRAFITY score and AFP in the fusion model may reflect absorption of their information by the imaging risk score,²⁰ whereas extrahepatic metastasis retained independent significance, suggesting that systemic dissemination encodes biology not captured by local perfusion imaging.

HepaKineticNet advances several methodological aspects. Conventional radiomics extracts features from static concatenated phases, discarding directional temporal information,^9,21 and prior deep learning models treat multi-phase stacks as static multi-channel inputs;^7,8 habitat-imaging approaches add subregion parcellation but lack dynamic-enhancement integration.¹⁷ The SpatialKDA module represents, to our knowledge, the first explicit incorporation of directional phase-to-phase kinetic differencing within a joint segmentation-heterogeneity network. Multiscale CV self-supervision enables heterogeneity mapping without prohibitive annotation. Conditioning radiomics on the heterogeneity map yields features with explicit geometric and statistical definitions, so that the final Cox model is fully transparent; the attention and ITH maps themselves, however, provide spatial saliency rather than mechanistic explanation, and do not by themselves render the deep learning component interpretable.²²

Discrimination improved with follow-up duration for both models, but the gap between them widened: in the external test cohort the time-dependent AUC of the fusion model rose from 0.760 at one year to 0.838 at two and 0.909 at three years, against 0.676, 0.742 and 0.772 for the clinical model, so that the absolute difference grew from 0.084 to 0.096 and 0.137. This pattern suggests that the prognostic information carried by ITH-derived spatial descriptors is weighted towards later events, which may reflect determinants of dissemination and acquired resistance rather than of early mortality. Both three-year estimates should nevertheless be read with caution. Only three patients remained at risk at 36 months in the external cohort, and the resulting uncertainty is severe: the confidence interval of the fusion model extends to unity, while that of the clinical model spans values below chance, indicating that the three-year landmark cannot support a reliable comparison between models in a cohort of this size, so the two-year values (0.838, 95% CI 0.736–0.930, vs 0.742) provide the more stable comparison.

Clinically, HepaKineticNet integrates into existing HCC workflows: it processes standard three-phase CECT acquired for staging, requires no additional imaging or biopsy, and outputs an interpretable continuous risk score. Utility across tumor size, multiplicity, vascular invasion, hepatic reserve, and etiology subgroups supports applicability in this combination-therapy population. Given the established link between imaging heterogeneity and immune microenvironment,^23,24 these features may serve as noninvasive surrogates for immune contexture to help guide ICI selection.

Several limitations merit acknowledgment. The population was restricted to patients receiving combined locoregional therapy with first-line tyrosine kinase inhibitors and immune checkpoint inhibitors, so generalizability to untreated or differently treated HCC remains to be established. The cohort was defined by treatment strategy rather than by stage, since this combination is applied across BCLC stages B and C in Chinese practice, and the model is intended to stratify precisely this group of candidates at the point of treatment decision. Extrahepatic metastasis, a defining criterion of stage C, remained independently prognostic alongside the imaging risk score in the multivariable model, which suggests that the two carry non-redundant information. Formal stage-specific model fitting was not attempted, because subdividing an external test cohort of 95 patients would leave too few events per stratum for stable estimation; possible heterogeneity of effect across stages therefore remains to be examined in larger cohorts.

Treatment allocation was not randomised and reflected drug availability, cost and reimbursement rather than tumour biology. Because no randomised evidence indicates a survival difference between the individual PD-1 inhibitors or between the available TKIs, and because subdivision by specific agent would leave fewer than ten events per stratum, agent-level modelling was not attempted. Residual confounding by treatment cannot be excluded, and the model should be interpreted as prognostic within this combination-therapy population rather than as predictive of benefit from any specific regimen.

HepaKineticNet is a baseline model and, like BCLC stage, ALBI grade and the CRAFITY score, characterises the tumour only at the pre-treatment time point; it cannot capture clonal evolution, on-treatment response or subsequent lines of therapy. Its intended use is therefore risk stratification at the moment of first treatment decision, and longitudinal delta-kinetic modelling across treatment cycles is the natural extension of this work.

The segmentation and heterogeneity-mapping network was trained by five-fold cross-validation on pooled images from both centres and therefore had prior exposure to the image distribution of the external centre, which may render the reported external performance slightly optimistic; no survival labels were used at that stage, so outcome information did not enter the prognostic model. Clinical variables were selected by LASSO-penalised Cox regression with 10-fold cross-validation, consistent with the radiomics pipeline, and the fusion model comprised four predictors at 32 events per variable; selection stability was nevertheless not assessed by resampling, so the retained set should not be regarded as uniquely determined. The median split is a pragmatic dichotomisation without clinical anchoring, and any threshold intended to guide patient care must be calibrated and validated prospectively. The retrospective bi-institutional design, despite consecutive enrollment, precludes causal inference and necessitates prospective external validation—particularly in Western populations with alcohol- and HCV-related HCC, where kinetic perfusion may differ from HBV-related disease. Scanner heterogeneity may introduce feature instability, warranting test-retest reproducibility studies. The analysis focuses on OS; future work should address progression-free survival, recurrence, and treatment response. Finally, the correspondence between HepaKineticNet hotspots and neoangiogenic or invasive-front regions is inferred from imaging physiology and from published radiogenomic work; it was not verified histologically in this retrospective cohort, in which most patients were diagnosed non-invasively and no spatially co-registered specimens were available. These associations are therefore hypothesis-generating and should not be read as biological evidence. A prospective sub-study is planned in which CT-guided biopsies with recorded sampling coordinates will be co-registered to HepaKineticNet hotspot maps and stained for CD34 and CD8, allowing the imaging–pathology correspondence proposed here to be tested directly.^25,26

Conclusion

In this bi-institutional retrospective study of BCLC stage B–C hepatocellular carcinoma treated with locoregional therapy plus first-line TKI and anti–PD-1 agents, kinetics-derived ITH maps carried prognostic information that modestly improved individualised OS prediction over clinical variables alone. Discrimination was maintained in an independent external test cohort from a second institution and across prespecified subgroups. Several limitations temper these findings: the risk threshold was prespecified as the median of the development cohort but has no clinical anchoring and has not been prospectively calibrated; treatment allocation was not randomised, so residual confounding cannot be excluded; and the proposed biological correlates of the imaging hotspots were not verified histologically. HepaKineticNet should therefore be regarded as a hypothesis-generating decision-support prototype rather than a tool ready for clinical use. Prospective evaluation is planned, including a sub-study in which CT-guided biopsies with recorded sampling coordinates will be co-registered to hotspot maps and stained for CD34 and CD8; prospective multicentre validation, together with clinical calibration of the risk threshold, will be required before the model can inform patient care. Practical implementation will additionally require standardised triphasic acquisition protocols, assessment of feature stability across scanners and vendors, and integration into routine reporting workflows.

Funding Statement

This research was supported by the Shanghai YRD Foundation for Innovation in Health Industry (2025-YRDFHI-012) and Project YXH2025YS059 supported by the Shandong Provincial Medical Association (Natural Science).

Abbreviations

AASLD, American Association for the Study of Liver Diseases; AFP, Alpha-fetoprotein; ALBI, Albumin–bilirubin;AP, Arterial phase; AUC, Area under the receiver operating characteristic curve; BCLC, Barcelona Clinic Liver Cancer; CECT, Contrast-enhanced computed tomography; CI, Confidence interval; C-index Concordance index; CRAFITY, C-reactive protein and alpha-fetoprotein in immunotherapy; CT, Computed tomography; cTACE, Conventional transarterial chemoembolization; DCA, Decision curve analysis; DP Delayed phase; DSA, Digital subtraction angiography; EASL, European Association for the Study of the Liver; ECOG PS, Eastern Cooperative Oncology Group performance status; FDR, False discovery rate; HAIC, Hepatic arterial infusion chemotherapy; HBV, Hepatitis B virus; HCC, Hepatocellular carcinoma; HCV, Hepatitis C virus; HR, Hazard ratio; IPCW, Inverse probability of censoring weighting; ITH, Intratumoral heterogeneity; NLR, Neutrophil-to-lymphocyte ratio; U-Net, Self-configuring U-Net-based segmentation framework; OS, Overall survival; PD-1, Programmed cell death protein 1; PLR, Platelet-to-lymphocyte ratio; PVP, Portal-venous phase; RECIST, Response Evaluation Criteria in Solid Tumors; ROC, Receiver operating characteristicSpatialK; DA, Spatial Kinetic Differential Attention; AUC, Time-dependent area under the curve; TKI, Tyrosine kinase inhibitor.

Data Sharing Statement

The datasets generated and analysed during the current study are available from the corresponding author, Dr Jinpeng Li (ljpxxx308@126.com), upon reasonable request.

Ethical Approval

This study was performed in accordance with the principles of the Declaration of Helsinki (as revised in 2013). It was approved by the Ethics Committee of Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences (approval no. SDTHEC202408020) and the Ethics Committee of Yantai Yuhuangding Hospital, Qingdao University (approval no. K2024-044). Owing to the retrospective design of the study, the requirement for written informed consent was waived by both ethics committees. All patient data were handled in strict confidence: records were de-identified before analysis, all direct identifiers were removed and replaced by study-specific codes, DICOM headers were anonymised, and data were stored on password-protected institutional servers accessible only to the investigators. No individually identifiable information is presented in this article or its supplementary material.

Consent to Participate

Written informed consent was waived because of the retrospective design of this study.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.