Abstract
Background
Although combining systemic and locoregional therapies has improved outcomes for hepatocellular carcinoma (HCC), identifying appropriate candidates remains challenging because no reliable tool predicts response.
Methods
This retrospective multicenter study enrolled 334 HCC patients who received first-line locoregional-systemic therapy from nine centers in China. Following automated tumor segmentation using nnU-Net on pretreatment contrast-enhanced magnetic resonance imaging, predefined radiomics features were computationally extracted, and deep learning features were obtained via a ResNet convolutional neural network. Three types of predictive models for tumor response (radiomics, deep learning, and combined) were developed and validated on the training cohort (n = 213), internal validation cohort (IVC, n = 53), and external validation cohort (EVC, n = 68). Performance was assessed using area under the receiver operating characteristic curve (AUC) and net reclassification improvement (NRI). Associations of model with progression-free survival (PFS) and overall survival (OS) were evaluated.
Results
Among the three models (multi-sequence radiomics, multi-sequence deep learning, and combined), the combined model showed the best predictive performance in both the training (AUC: 0.997, 95% CI: 0.994–1.000) and validation cohorts (AUC of IVC 0.900, 95% CI: 0.806–0.994; AUC of EVC: 0.871, 95% CI: 0.790–0.951). The combined model demonstrated significantly improved predictive performance compared with the multi-sequence radiomic model (training NRI = 43.9%, p < 0.001; IVC NRI = 38.7%, p = 0.038; EVC NRI = 3.2%, p = 0.806) and multi-sequence deep learning model (training NRI = 80.8%, IVC NRI = 72.7%, EVC NRI = 50.4%; all p < 0.05). Specifically, the combined model was significantly associated with both PFS and OS.
Conclusions
The combined radiomics-deep learning model outperformed both single-modality models in predicting objective response and stratifying survival outcomes in unresectable HCC patients receiving combined systemic and locoregional therapy.
Keywords: Hepatocellular carcinoma, Systemic therapy, Locoregional therapy, Radiomics, Deep learning
Introduction
Hepatocellular carcinoma (HCC) is the sixth most common cancer worldwide and ranks third in cancer-related mortality [1, 2]. The prognosis for HCC remains poor, primarily because more than 70% of patients present with unresectable disease at diagnosis, thereby missing the window for potentially curative treatments [3].
With the rapid development and diversification of treatment options for unresectable HCC in recent years, combined systemic and locoregional therapies have demonstrated superior clinical outcomes compared with systemic therapy alone [4, 5] or locoregional therapy alone [6–9] in patients with intermediate-to-advanced HCC, particularly regarding short-term outcomes, including objective response rate (ORR) and progression-free survival (PFS). However, the concomitant increase in treatment-related adverse events (AEs) associated with the combination therapy underscores the need for predictive biomarkers to identify patients who may derive clinical benefit from combination therapy. Specifically, without a predictive tool, patients who are unlikely to respond to combination therapy would be unnecessarily exposed to higher toxicity. Therefore, the development of a predictive model that estimates individual benefit is a logical and essential step toward precision combination therapy, and predicting tumor response to combination therapy remains critical for the management of unresectable HCC.
Recent advances in noninvasive imaging have shown promising potential for predicting treatment responses in HCC patients receiving various therapeutic approaches. Studies have proposed various biomarkers, ranging from magnetic resonance imaging (MRI)-based features for immunotherapy [10] and radiomic nomograms for locoregional therapy [11] to CT-based models for combination therapy with lenvatinib plus programmed cell death 1 (PD-1) inhibitors and interventional procedures [12]. Radiomics extracts high-throughput quantitative features to decode the subvisual tumor characteristics [13], yet it remains constrained by reliance on engineered features and segmentation accuracy. In contrast, deep learning demonstrates strong performance in medical imaging through end-to-end feature learning, but its limited interpretability and need for large annotated datasets pose clinical translation challenges [14]. A growing body of research suggests that integrating these approaches is a promising direction for enhancing predictive models in medical artificial intelligence [15–19]. Nevertheless, in predicting treatment response for HCC patients receiving combination therapy, whether such integration offers superior performance in addressing this multifactorial clinical problem remains insufficiently explored.
This retrospective study conducted a comparative analysis between radiomics-based and deep learning-based predictive models for tumor response in unresectable HCC patients receiving combination therapy. Based on this comparison, we systematically examined whether a hybrid model integrating both approaches demonstrates enhanced performance and robustness compared to either method alone, with the ultimate goal of supporting pretreatment clinical decision-making.
Methods
This retrospective, multicenter study was conducted in 9 tertiary care referral hospitals in China. The study protocol was approved by the Zhongshan Hospital Research Ethics Committee (Approval No. B2021-210(5)). This study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all patients before enrollment.
Study Design and Patients
The study workflow is outlined in Figure 1. Between September 2018 and January 2025, medical records and MRI imaging data from consecutive patients who fulfilled the following inclusion criteria were collected: (a) age ≥18 years; (b) with clinically diagnosed (based on the American Association for the Study of Liver Disease criteria) or pathologically confirmed HCC [20, 21]; (c) received first-line systemic therapy combined with locoregional therapy, defined as two therapies administered within a 1-month interval, irrespective of the sequence; (d) underwent contrast-enhanced MRI within 4 weeks before initiation of combined therapy; (e) had at least one measurable untreated lesion in the liver per modified Response Evaluation Criteria in Solid Tumors (mRECIST); and (f) had tumor response assessed every 2 months (±2 weeks) post-treatment via contrast-enhanced CT or MRI according to mRECIST. Patients were excluded if they (a) had any previous or current malignancy other than HCC; (b) had inadequate MRI quality for analyses (e.g., incomplete images, severe artifacts); (c) had no identifiable target lesions in the liver; and (d) lacked any follow-up information. The participating centers and the corresponding number of patients in each cohort are detailed in the online supplementary Methods (for all online suppl. material, see https://doi.org/10.1159/lic/admag009).
Fig. 1.
Flowchart for this study.
The Combination Therapy
All patients received first-line systemic therapy with targeted therapy and immunotherapy, along with locoregional therapy. The specific treatment regimen was determined by experienced clinicians based on the patient’s physical condition and in consultation with the patient. Details of the systemic treatment regimens (including dosages) and locoregional therapy modalities are described in the online supplementary Methods.
Follow-Up and Endpoint
Patients were followed every 2 months (±2 weeks) after initiation of combination therapy. Overall survival (OS) was calculated from the date of combination therapy to death from any cause or was censored on the last follow-up. PFS was calculated from the date of combination therapy to the first documented disease progression or death from any cause. The primary endpoint was the confirmed objective response (OR), defined as the best overall response of either complete response or partial response based on mRECIST criteria [22], with the response confirmed by first two consecutive tumor response assessments (performed at protocol-specified intervals, typically every 2 months [±2 weeks]).
Evaluation of Clinical Features
Clinical features with p < 0.05 in multivariable logistic regression were deemed independently associated. A prediction model would be constructed by incorporating independent clinical features.
Pipeline of Prediction Model Development and Validation
The model development pipeline followed a structured approach. First, multicenter data were pooled and randomly split into training and internal validation cohorts (IVCs) (8:2 ratio). Following nnU-Net-based auto-segmentation, both radiomic features and ResNet-derived deep features were extracted. Second, using the training data, three distinct predictive models (radiomics, deep learning, and a combined model) were constructed. Third, a two-tier validation strategy was employed, first through IVC and then on an external validation cohort (EVC) comprised temporally independent data from two of the participating centers, to rigorously assess model generalizability.
MRI Acquisition and Automatic Tumor Segmentation
MRI scans were performed using either 1.5-T or 3.0-T systems, with either extracellular or hepatobiliary contrast agents administered as appropriate. The imaging protocol followed standard liver MRI sequences. Further details on the analyzed sequences are provided in the online supplementary Methods.
This study utilized automated tumor segmentation on MRI images based on the deep learning framework nnU-Net [23]. The encoder progressively reduced the spatial dimensions of the image, while the decoder gradually restored these dimensions. Details of the nnU-Net network are presented in the online supplementary Methods. Evaluation on a test cohort of 54 patients yielded excellent performance, demonstrating a 100% correct sequence identification rate and a mean Dice coefficient of 0.889 for region-of-interest segmentation. All automated segmentation results underwent qualitative validation by two board-certified radiologists.
Radiomic Model
To mitigate potential confounding effects from MRI data from 9 centers, data resampling standardization was performed. To standardize voxel distribution across multicenter images, voxel spacing was normalized to an isotropic resolution using nearest neighbor interpolation. This method preserves original image information while providing a uniform foundation for subsequent radiomics feature extraction. Imaging features were extracted from each of the seven MRI sequences using PyRadiomics (v3.0.1) [24]. The seven sequences were arterial phase, portal venous phase, delayed phase, T2-weighted imaging, diffusion-weighted imaging, T1-weighted chemical shift in-phase and out-of-phase (T1 CS IP and OP). Features extracted for each image type are detailed in online supplementary Table S1. Finally, 1,118 features were extracted for each of the seven sequences (7,826 features in total).
Subsequently, Z-score normalization was applied to the radiomic features. For feature selection in the training cohort, we employed a dual approach combining minimum redundancy maximum relevance [25] and least absolute shrinkage and selection operator [26] methods to mitigate bias and overfitting risks. These complementary methods evaluate features from different perspectives. Minimum redundancy maximum relevance assesses feature relevance while maintaining diversity, and least absolute shrinkage and selection operator performs regularization-based selection, thereby comprehensively considering both predictive importance and correlation. The selected features were then used to construct radiomic models based on decision tree, logistic regression, random forest, and support vector machine, with separate models developed for each imaging sequence. For each sequence, the algorithm that achieved the best performance on the IVC was selected as the final result. Radiomic model development was conducted on the United Imaging uAI Research Portal platform [27]. The predicted probabilities from each single-sequence model were integrated using a logistic regression to build a multi-sequence radiomic model.
Deep Learning Model
Deep learning features were learned from pretreatment MRI using both 2.5D and 3D ResNet models. For thin-slice phase images, both 2.5D and 3D deep learning frameworks were implemented, ultimately yielding ten distinct deep learning models. The specific process is detailed in the online supplementary Methods.
Model training was conducted using PyTorch’s DataLoader with a batch size of 32 (Version 2.0.1; https://pytorch.org), incorporating real-time data augmentation including random grayscale transformation, rotation, flipping, and scaling to enhance data diversity. We employed the ASLSingleLabel loss function (smoothing factor = 0.1, γ_pos = 0, γ_neg = 4) and optimized the model using AdamW (initial learning rate = 1e−5) with Warmup Cosine scheduling. The ResNet34 [28]-based training protocol included periodic weight checkpointing and loss monitoring to ensure model convergence. Single-sequence deep learning models were built for all included sequences; their prediction probabilities were then integrated using logistic regression to develop a multi-sequence deep learning model.
Combined Model
The combined model was developed through logistic regression fusion of the output probabilities from both the multi-sequence radiomic model and the multi-sequence deep learning model.
Model Performance
The predictive performance of the models was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (estimated using 1,000 stratified bootstrap replicates), accuracy, sensitivity, specificity, precision, and F1 score. The optimal cutoff threshold for the prediction models was determined using ROC analysis by maximizing the Youden index in the training cohort. Calibration curves were plotted to assess model calibration. Decision curve analysis was conducted to determine the clinical utility of the prediction models by quantifying the net benefits at different threshold values. The net reclassification index (NRI) was used to compare the discriminatory abilities between two models.
Statistical Analysis
Categorical data are presented as frequencies and proportions, with group comparisons performed using Pearson’s χ2 test, Fisher’s exact test, or the Mann-Whitney U test, depending on data characteristics. Continuous variables are summarized as mean (± standard deviation) or median (interquartile range [IQR]), with between-group differences assessed via Student’s t test or the Mann-Whitney U test, as appropriate. Survival analyses were conducted using Kaplan-Meier estimates, with between-group differences evaluated by the log-rank test. All statistical tests were two-tailed, with a significance threshold set at p < 0.05. Analyses were performed using R (version 4.3.1) and Python (version 3.10).
Results
Patients Characteristics
Figure 2 illustrates the patient inclusion and exclusion process. A total of 334 eligible patients were enrolled and allocated as follows: 213 in the training cohort and 53 in IVC based on a random 8:2 ratio, plus an additional 68 in EVC. Significant baseline differences were observed among these cohorts (Table 1). Specifically, the EVC demonstrated higher rates of both bevacizumab-based targeted therapy and HAIC plus TACE administration, as well as elevated baseline protein induced by vitamin K absence or antagonist-II levels, compared to the training cohort and the IVC. The ORRs per mRECIST were comparable across three cohorts: 57.3% (122/213) in the training cohort, 56.6% (30/53) in the IVC, and 58.8% (40/68) in the EVC (p = 0.965).
Fig. 2.
Flowchart of patient enrollment.
Table 1.
Baseline characteristics of patients in the training and validation cohorts
| | Training cohort (n = 213) | IVC (n = 53) | EVC (n = 68) | p value |
|---|---|---|---|---|
| Sex, n (%) | | | | 0.002 |
| Female | 37 (17.4) | 11 (20.8) | 1(1.5) | |
| Male | 176 (82.6) | 42 (79.2) | 67 (98.5) | |
| Age, years | 56.7 (11.5) | 56.5 (10.5) | 53.5 (11.3) | 0.127 |
| Targeted therapeutic agents, n (%) | | | | <0.001 |
| Bevacizumab (biosimilar) | 74 (34.7) | 24 (45.3) | 65 (95.6) | |
| TKI | 139 (65.3) | 29 (54.7) | 3 (4.4) | |
| Types of local treatment, n (%) | | | | <0.001 |
| HAIC | 83 (39.0) | 16 (30.2) | 19 (27.9) | |
| TACE | 117 (54.9) | 32 (60.4) | 12 (17.7) | |
| TACE + HAIC | 13 (6.10) | 5 (9.4) | 37 (54.4) | |
| ECOG PS, n (%) | | | | 0.079 |
| 0 | 155 (72.8) | 39 (73.6) | 63 (92.7) | |
| 1 | 22 (10.3) | 5 (9.4) | 2 (2.9) | |
| NA | 36 (16.9) | 9 (17.0) | 3 (4.4) | |
| Child-Pugh, n (%) | | | | 0.539 |
| A | 185 (86.9) | 45 (84.9) | 68 (100.0) | |
| B | 4 (1.9) | 1 (1.9) | 0 (0.0) | |
| NA | 24 (11.3) | 7 (13.2) | 0 (0.0) | |
| CNLC stage, n (%) | | | | 0.158 |
| Ia | 6 (2.82) | 1 (1.89) | 0 (0.00) | |
| Ib | 28 (13.15) | 11 (20.75) | 12 (17.65) | |
| IIa | 10 (4.69) | 3 (5.66) | 0 (0.00) | |
| IIb | 38 (17.84) | 9 (16.98) | 13 (19.12) | |
| IIIa | 102 (47.89) | 24 (45.28) | 26 (38.24) | |
| IIIb | 29 (13.62) | 5 (9.43) | 17 (25.00) | |
| BCLC stage, n (%) | | | | 0.774 |
| A | 34 (16.0) | 12 (22.6) | 12 (17.7) | |
| B | 48 (22.5) | 12 (22.6) | 13 (19.1) | |
| C | 131 (61.5) | 29 (54.7) | 43 (63.2) | |
| HBsAg, n (%) | | | | 0.638 |
| Negative | 33 (15.5) | 8 (15.1) | 8 (11.8) | |
| Positive | 170 (79.8) | 39 (73.6) | 60 (88.2) | |
| NA | 10 (4.7) | 6 (11.3) | 0 (0.0) | |
| HBV DNA, n (%) | | | | 0.210 |
| ≤103 IU/mL | 104 (48.8) | 18 (34.0) | 38 (55.89) | |
| >103 IU/mL | 85 (40.0) | 26 (49.1) | 30 (44.1) | |
| NA | 24 (11.3) | 9 (17.0) | 0 (0.0) | |
| AFP, ng/mL | 163.50 [16.42; 2,588.00] | 34.20 [4.20; 3,696.00] | 595.50 [16.45; 17,038.00] | 0.185 |
| AFP, n (%) | | | | 0.364 |
| ≤400 ng/mL | 123 (57.8) | 31 (58.5) | 33 (48.5) | |
| >400 ng/mL | 89 (41.8) | 22 (41.5) | 35 (51.5) | |
| NA | 1 (0.4) | 0 (0.0) | 0 (0.0) | |
| PIVKA-II, mAU/mL | 3,143.00 [732.20; 20,948.75] | 13,305.50 [658.84; 50,825.75] | 26,003.00 [3,641.50; 30,000.00] | 0.015 |
| PIVKA-II, n (%) | | | | 0.132 |
| ≤1,000 mAU/mL | 60 (28.2) | 13 (24.5) | 12 (17.7) | |
| >1,000 mAU/mL | 136 (63.8) | 33 (62.3) | 55 (80.8) | |
| NA | 17 (8.0) | 7 (13.2) | 1 (1.5) | |
| BOR per mRECIST, n (%) | ||||
| CR | 23 (10.8) | 7 (13.2) | 19 (27.9) | 0.011 |
| PR | 99 (46.5) | 23 (43.4) | 21 (30.9) | |
| SD | 71 (33.3) | 19 (35.8) | 26 (38.2) | |
| PD | 20 (9.4) | 4 (7.6) | 2 (2.9) | |
| OR per mRECIST, n (%) | | | | 0.965 |
| No | 91 (42.7) | 23 (43.4) | 28 (41.2) | |
| Yes | 122 (57.3) | 30 (56.6) | 40 (58.8) | |
AFP, α-fetoprotein; BCLC, Barcelona Clinic Liver Cancer; BOR, best objective response; CNLC, China Liver Cancer; CR, complete response; ECOG, PS Eastern Cooperative Oncology Group performance status; HAIC, hepatic artery infusion chemotherapy; HBsAg, hepatitis B surface antigen; HBV DNA, hepatitis B virus DNA copy number; mRECIST, modified Response Evaluation Criteria in Solid Tumors; NA, not available; PD, progressive disease; PIVKA-II, protein induced by vitamin K absence or antagonist-II; PR, partial response; SD, stable disease; TACE, transarterial chemoembolization; TKI, tyrosine kinase inhibitor.
The baseline demographic and disease characteristics of patients with and without an OR in the training and validation cohorts are summarized in online supplementary Tables S2–4. In the training cohort, compared with the OR group, the non-OR population had a higher proportion of patients with Child-Pugh class A (89.0% vs. 85.2%, p = 0.039). In the IVC, the OR group featured more patients with the Eastern Cooperative Oncology Group performance status (ECOG PS) 0 (89.0% vs. 85.2%, p = 0.006) and a greater proportion with HBV DNA <1,000 IU/mL (46.7% vs. 17.4%, p = 0.043), compared with the non-OR population. In the EVC, patients with OR had an earlier BCLC stage (p = 0.006).
Clinical Features Were Not Associated with Tumor Response
Assessment of clinical features through logistic regression demonstrated that none of the baseline characteristics showed statistically significant correlation with OR in the training cohort (online suppl. Table S5).
Correlation between Radiomic Features and Tumor Response
After feature selection, the radiomic features from each sequence that were ultimately included in the model are shown in Table S6. Initially, we developed individual radiomic models for each of the 7 sequences. The AUC (95% CI), accuracy, sensitivity, specificity, precision, and F1 score of the training cohort and IVC for the seven models are summarized in online supplementary Table S7. Among all phases analyzed, the in-phase (IP) phase radiomic model demonstrated the best predictive performance, with AUC values (95% CI) of 0.977 (0.961–0.992) in the training cohort and 0.736 (0.598–0.874) in the IVC (online suppl. Table S7). Given that the predictive performance of the single-sequence radiomic model was relatively modest, we subsequently developed a comprehensive model integrating all seven sequences.
The radiomic model incorporating all imaging sequences demonstrated IVC AUC (95% CI) of 0.786 (0.656–0.916) (Fig. 3a). The radiomic model integrating all sequences showed a higher AUC in the IVC than the IP phase single-sequence model (0.786 vs. 0.736), though this difference was not statistically significant (p = 0.249). However, a significant and substantial NRI was observed for the integrated model in both the training cohort (59.8%, p < 0.001) and IVC (56.1%, p = 0.017). This suggests that integrating multiple sequences enhances the predictive power of the radiomic model for treatment efficacy classification compared with a single-sequence approach. Subsequently, this integrated model was validated in the EVC, achieving an AUC (95% CI) of 0.849 (0.762–0.936) (Fig. 3b). The optimal cutoff threshold for this integrated model was 0.500. The AUC (95% CI), accuracy, sensitivity, specificity, precision, and F1 score of the training and validation cohorts for the multi-sequence radiomic model are summarized in Table 2. As shown in online supplementary Figure S1a, b, the confusion matrix for the validation cohorts demonstrated the model’s performance. Furthermore, the calibration curves (online suppl. Fig. S2a, b) for the validation cohorts indicated good agreement between the predictions made by the multi-sequence radiomic model and the actual observations.
Fig. 3.
Performance of the radiomics, deep learning, and combined models. a, b The ROC of the multi-sequence radiomics, the multi-sequence deep learning, and combined models in the IVC and EVC. c, d Confusion matrix of combined model in the IVC and EVC. e, f The scores predicted by three models between patients with and without OR in the IVC and EVC. g, h The decision curve analysis of three models in the IVC and EVC.
Table 2.
Performance of predictive models
| | AUC (95% CI) | Sensitivity | Specificity | Accuracy | Precision | F1 score |
|---|---|---|---|---|---|---|
| The multi-sequence radiomic model | ||||||
| Training cohort | 0.927 (0.895–0.959) | 0.828 | 0.868 | 0.845 | 0.894 | 0.860 |
| IVC | 0.786 (0.656–0.916) | 0.600 | 0.696 | 0.642 | 0.720 | 0.655 |
| EVC | 0.849 (0.762–0.936) | 0.700 | 0.750 | 0.721 | 0.800 | 0.747 |
| The multi-sequence deep learning model | ||||||
| Training cohort | 1.000 (1.000–1.000) | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 |
| IVC | 0.829 (0.714–0.944) | 0.767 | 0.826 | 0.793 | 0.852 | 0.807 |
| EVC | 0.740 (0.615–0.865) | 0.275 | 0.964 | 0.559 | 0.917 | 0.421 |
| Combined model | ||||||
| Training cohort | 0.997 (0.994–1.000) | 0.967 | 0.978 | 0.972 | 0.983 | 0.975 |
| IVC | 0.900 (0.806–0.994) | 0.867 | 0.739 | 0.811 | 0.813 | 0.839 |
| EVC | 0.871 (0.790–0.951) | 0.700 | 0.750 | 0.721 | 0.800 | 0.747 |
AUC, area under the receiver operating characteristic curve; CI, confidence interval.
Correlation between Deep Learning Features and Tumor Response
Deep learning feature extraction and model development were performed separately across the seven MRI sequences, employing both 2.5D and 3D frameworks, which ultimately resulted in ten single-sequence deep learning models. The AUC (95% CI), accuracy, sensitivity, specificity, precision, and F1 score of the training cohort and IVC for the ten models are summarized in online supplementary Table S8. Among the ten independently developed deep learning models for individual sequences, most demonstrated modest predictive capability. The 3D-delay phase model emerged as the top-performing single-sequence model, achieving excellent discrimination in the training cohort (AUC = 1.000 [95% CI: 1.000–1.000]) and moderate discrimination in the IVC (AUC = 0.771 [95% CI: 0.706–0.837]). Similarly, recognizing the suboptimal predictive performance of the single-sequence deep learning model, we subsequently constructed a composite model that integrated all seven sequences.
The multi-sequence deep learning model integrating all ten single-sequence models achieved AUC [95% CI] of 0.829 [0.714–0.944] in the IVC) (Fig. 3a). Although the deep learning model incorporating all sequences did not achieve a statistically significant increase in the IVC AUC compared with the 3D-delay phase model (0.829 vs. 0.771, p = 0.207), it yielded a marked improvement in predictive performance as measured NRI. The NRI was significant and substantial for both the training cohort (121.8%, p < 0.001) and IVC (75.9%, p = 0.002). This finding also suggests that the multi-sequence deep learning model possesses stronger discriminatory capacity for predicting therapeutic efficacy than the single-sequence model. We then validated this integrated model in the EVC, which yielded an AUC of 0.740 (95% CI: 0.615–0.865) (Fig. 3b). The optimal cutoff for this model was 0.630. The AUC (95% CI), accuracy, sensitivity, specificity, precision, and F1 score of the training and validation cohorts for the multi-sequence deep learning model are summarized in Table 2. Model performance on the validation cohorts was evaluated by the confusion matrix (online suppl. Fig. S1c, d) and further evaluated using calibration curves (online suppl. Fig. S2c, d), which showed good agreement between predictions and observations.
The Combined Model to Predict Tumor Response
Although the multi-sequence radiomics and deep learning models generally showed superior performance, their validation results in the IVC and EVC were unstable (Table 2). We therefore explored whether integrating these two approaches could further improve performance. As shown in Figure 3a, b and Table 2, the combined model achieved higher AUCs in both validation cohorts compared with either individual model. NRI analysis further demonstrated that the combined model significantly outperformed the multi-sequence radiomic model in the training cohort and the IVC (training NRI = 43.9%, p < 0.001; IVC NRI = 38.7%, p = 0.038), whereas no significant difference was observed in the EVC (NRI = 3.2%, p = 0.806). Moreover, the combined model also showed substantial improvement over the multi-sequence deep learning model (training NRI = 80.8%, p < 0.001; IVC NRI = 72.7%, p < 0.001; EVC NRI = 50.4%, p = 0.016). The accuracy, sensitivity, specificity, precision, and F1 score of the training and validation cohorts for the combined model are summarized in Table 2. As illustrated in Figure 3c, d, the confusion matrix for the validation cohorts detailed the classification results. The calibration curves for the validation cohorts indicated good agreement between the combined model’s predictions and actual observations (online suppl. Fig. S2e, f).
In the validation cohorts, the prediction scores from the multi-sequence radiomic model, the multi-sequence deep learning model, and the combined model all revealed significant differences between patients predicted as OR and those predicted as non-OR (Fig. 3e, f). Decision curve analysis demonstrated superior net clinical benefit of the combined model over the multi-sequence radiomics and multi-sequence deep learning models across all threshold probabilities in the validation cohorts (Fig. 3g, h). These results highlighted that the combined model surpasses both the multi-sequence radiomic model and the multi-sequence deep learning model in predictive performance.
Subgroup Analysis of Combined Model Performance
Using the combined model, we subsequently conducted subgroup analyses in the validation cohorts based on systemic therapy, locoregional therapy, tumor stage, and baseline AFP level. Patients were stratified according to systemic treatment regimens, bevacizumab-based or TKI-based therapy; locoregional treatment modalities, TACE alone, HAIC alone, or TACE+HAIC; tumor staging per BCLC classification system, BCLC-A or BCLC-B–C subgroups; and pretreatment AFP levels using the 400 ng/mL cutoff, ≥400 versus <400 ng/mL. The model demonstrated robust performance across all subgroups, with the performance in the BCLC-A subgroup being slightly lower and less stable, likely due to the small number of patients in this category (Fig. 4).
Fig. 4.
Subgroup analysis of combined model.
Predicting Survival with the Radiomic Model, Deep Learning Model, and Combined Model
As of October 20, 2025, the median follow-up time was 14.30 (IQR: 9.73–20.20) months in the training cohort, 11.83 (IQR: 6.87–21.00) months in the IVC and 18.90 (IQR: 8.18–26.74) in the EVC. The median OS was 43.47 (95% CI: 32.20–NE) months and 29.77 (95% CI: 21.40–NE) months in the training cohort and IVC, and the median OS was not reached in the EVC (online suppl. Fig. S3a). The median PFS was 9.63 (95% CI: 8.37–11.20) months in the training cohort, 9.23 (95% CI: 7.97–15.60) months in the IVC, and 25.17 (95% CI: 18.13–NE) months in the EVC (online suppl. Fig. S3b). In the training cohort, both the multi-sequence deep learning (online suppl. Fig. S4a, b) and combined models (Fig. 5a, b) demonstrated significant correlations between prediction scores and OS/PFS, while the multi-sequence radiomic model’s prediction scores showed a correlation with PFS alone (online suppl. Fig. S5a, b). In the IVC, the prediction scores of both the multi-sequence deep learning model and the multi-sequence deep radiomic model showed no significant correlation with OS or PFS (online suppl. Fig. S4c, d, S5c, d). By contrast, the prediction scores of the combined model were correlated with both OS (p = 0.052) and PFS (p = 0.046) (Fig. 5c, d). In the EVC, however, the multi-sequence radiomic model showed significant correlations with both OS (p = 0.009) and PFS (p = 0.008) (online suppl. Fig. S5e, f). The multi-sequence deep learning model remained partially effective, demonstrating a correlation only with PFS (p = 0.045) (online suppl. Fig. S4e, f). Notably, the combined model consistently exhibited significant correlations with both OS (p = 0.009) and PFS (p = 0.008) in this cohort (Fig. 5e, f).
Fig. 5.
Survival associations with combined model. a, b Association between combined model and OS/PFS in the training cohort. c, d Association between combined model and OS/PFS in the IVC. e, f Association between combined model and OS/PFS in the EVC.
Discussion
This study developed radiomics, deep learning, and combined radiomics-deep learning models based on pretreatment MRI to predict tumor response to combination treatment with locoregional therapy and systemic therapy, and compared their predictive performance, with the aim of providing a tool for treatment selection in unresectable HCC patients. The results indicated that both the multi-sequence radiomic model and the deep learning model demonstrated superior predictive performance compared with the single-sequence model. By integrating radiomics and deep learning approaches, predictive capability was further enhanced with more stable performance across validation cohorts, and the combined model showed a significant association with survival outcomes.
Although recent studies of combined systemic and locoregional therapy have employed a variety of drugs and TACE regimens, all three reported ORRs consistently fall within a narrow range of 40–50% [6, 7, 29]. This highlights a critical unmet need for predictive tools to identify responders. The present study demonstrated that the combined radiomics-deep learning model achieved optimal predictive performance in both the training (AUC: 0.997) and validation (AUC: 0.900 and 0.871) cohorts, indicating excellent discrimination and robustness. This combined model demonstrated significantly superior reclassification ability compared to both the multi-sequence radiomic model (IVC NRI = 38.7%, p = 0.038; EVC NRI = 3.2%, p = 0.806) and the multi-sequence deep learning model (IVC NRI = 72.7%; EVC NRI = 50.4%, all p < 0.05). The superior results underscore the complementary value of multimodal fusion, where predefined radiomic features and automatically learned deep representations synergistically contribute distinct predictive information, thereby substantially enhancing the model's accuracy and clinical utility for complex medical image analysis.
Multiple serum and tumor molecular biomarkers have demonstrated prognostic value for HCC patients receiving systemic therapies or locoregional treatments [30–33]. However, these biomarkers require invasive sampling, while imaging biomarkers offer distinct advantages due to their noninvasiveness, ability to evaluate the entire tumor, and suitability for dynamic monitoring, making them highly translatable for clinical use.
Current research on imaging-based prediction of HCC treatment response has predominantly focused on CT radiomics approaches. For example, Vithayathil et al. [34] demonstrated that radiomic-based machine learning models could effectively stratify risk in patients with advanced HCC undergoing combination therapy with atezolizumab and bevacizumab. Similarly, Bo et al. [35] developed CT-based machine learning radiomic models that successfully predicted treatment response to lenvatinib monotherapy in patients with unresectable HCC. However, in medical imaging research for HCC, MRI offers significant advantages over CT owing to its multi-sequence capabilities and superior tissue resolution. Accordingly, this study utilizes multi-parametric MRI, which can capture treatment response-related features more comprehensively compared than CT radiomics.
While radiomics and deep learning each offer transformative potential for predicting HCC treatment response and personalizing care [14, 36–45], current research remains constrained. Most studies have employed unimodal approaches – using either radiomics or deep learning in isolation – and have focused on predicting outcomes for a single therapeutic regimen, thereby failing to harness the full power of integrated methodologies. For instance, Xu et al. [46] developed a pretreatment MRI-based radiomic model to predict individual responses to lenvatinib plus anti-PD-1 antibody combination therapy in advanced HCC. Similarly, Liao et al. [47] employed a deep learning approach, integrating multiphase CT with algorithm-generated heatmaps to guide this same treatment regimen. While demonstrating clinical utility, these unimodal approaches may not fully capture the complex tumor response patterns that can be identified through combined feature analysis. This study represents the first application of a multimodal model in this context, which amalgamated radiomics with deep learning to enable more comprehensive extraction of imaging information and enhanced predictive capability for HCC patients undergoing combination therapy. Moreover, our study included patients with diverse systemic therapies, various locoregional treatment regimens, and different tumor stages. Despite multiple confounding factors, the combined model demonstrated robust and stable performance, which may be attributed to the synergistic integration of dual-dimensional imaging analysis approaches.
According to the model validation results, among patients predicted to achieve OR, the actual probability of achieving OR exceeded 80%. Therefore, for patients predicted to respond, combination therapy should be actively pursued under the conditions of close monitoring and management of AEs. For patients predicted not to achieve OR, treatment intensity can be adjusted, especially when AEs occur. Furthermore, among patients predicted not to achieve OR, there is a 9.3% probability of developing progressive disease (PD); thus, the decision to use combination therapy should be made with caution. However, the present study did not use PD as a prediction endpoint, mainly due to the low incidence of PD. In the future, with a sufficient sample size, PD should be considered as a clinical endpoint. Importantly, a predictive tool that identifies patients unlikely to achieve OR (predicted nonresponders) can directly inform clinical decision-making. For such patients, alternative strategies should be considered rather than proceeding with the standard combination regimen. These alternatives may include: (i) switching to a different systemic agent (e.g., alternative tyrosine kinase inhibitors or immune checkpoint inhibitors) that might bypass primary resistance mechanisms; (ii) locoregional monotherapy (e.g., TACE alone, radiofrequency ablation, or radiotherapy) for patients who still could benefit from liver-directed treatments while avoiding unnecessary systemic toxicity; (iii) enrollment in clinical trials evaluating novel agents or novel combination paradigms, given that predicted nonresponders represent an underserved population with high unmet need; or (iv) in the future, developing models capable of predicting the efficacy of different specific regimens to enable more precise treatment recommendations. Incorporating such decision pathways into future prospective studies will be essential to translate model predictions into improved patient outcomes.
Several limitations should be acknowledged in this study. First, the relatively small sample size may limit statistical power. Second, the retrospective design introduces inherent risks of bias. Third, the absence of model interpretability analysis might hinder clinical acceptance. The proposed model is not ready for standalone clinical decision-making without prospective validation and additional interpretability tools. To manage associated risks, we propose that any future clinical application would require: (i) external validation in multicenter prospective cohorts, (ii) integration with clinician oversight (human-in-the-loop), and (iii) clear communication to end users that the model provides risk stratification, not diagnosis or treatment prescription. Fourth, while the integration of radiomics and deep learning features from multiple dimensions enhanced model performance, it also increased the computational burden for clinical implementation. Finally, model training was affected by treatment heterogeneity, a confounding factor intrinsic to the multi-treatment prediction task. Nevertheless, the overall findings suggest that patients who respond to combination therapy may share common characteristics, despite treatment variety. To address these limitations, future studies should focus on expanding cohort sizes through multicenter collaborations to improve generalizability, investigating temporal imaging patterns to capture dynamic treatment responses, and optimizing model architectures for clinical deployment. Such improvements will enhance prediction accuracy and facilitate translation into routine practice, ultimately supporting personalized treatment decisions for HCC patients.
Conclusion
The combined radiomics-deep learning model demonstrated superior and stable predictive performance compared with either the radiomic model or the deep learning model alone in patients receiving combined treatment using systemic therapy and locoregional therapies, with consistently strong predictive capabilities across various subgroups.
Acknowledgments
The authors wish to thank all the patients and their families. We are also grateful to Mr. Hong-Tao Li for his valuable assistance in collecting the data from all participating centers.
Statement of Ethics
The study protocol was approved by the Zhongshan Hospital Research Ethics Committee (Approval No. B2021-210(5)). This study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all patients before enrollment.
Conflict of Interest Statement
H.-C. Sun received honorariums and lecture fees from Roche, Bayer, MSD, Eisai, Hengrui, Innovent, TopAlliance, Abbott, BeiGene, Gilead, and Zelgen over the last 5 years. The remaining authors have no conflicts of interest to declare.
Funding Sources
This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0520400 and 2024ZD0520401 to H.-C. Sun), the National Natural Science Foundation of China (82372037 to H-C Sun and 82503971 to B. Xu), the Special Research Fund for Liver Cancer Diagnosis and Treatment from the China Anti-Cancer Association (H2020-008 to H.-C. Sun), the Clinical Research Special Fund of Zhongshan Hospital, Fudan University (2020ZSLC71 to H.-C. Sun), the 2022 Personalized Medical Incubator Project, through the fund for Precision Medicine Research and Industry Development in SIMQ (EH20 to H.-C. Sun), and Outstanding Resident Clinical Postdoctoral Program of Zhongshan Hospital Affiliated to Fudan University (2025ZYYS-049 to B. Xu).
Author Contributions
L.-N.W. and S.-Q.Z. conceived and designed the study, analyzed the data, and drafted and revised the manuscript. F.L. and L.-F.W. were responsible for image collection and preprocessing. H.-C.S. and B.X. designed the study, interpreted the results, and revised the manuscript. L.X., K.W., N.-L.G., Y.C., T.P., T.-Q.S., M.-L.Y., L.Z., L.-D.Z., Y.-J.C., G.-D.Z., Y.-X.P., H.X., G.-Z.Z., D.-M.L., Z.-X.Z., K.H., Y.-C.Y., Z.-Y.Y., T.H., Z.-Y.W., C.H., Y.-H.S., X.-D.Z., H.L., Q.G., J.Z., and J.F. were responsible for the collection of patient clinical and imaging data. All authors provided critical comments regarding the manuscript and have read and approved this manuscript.
Funding Statement
This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0520400 and 2024ZD0520401 to H.-C. Sun), the National Natural Science Foundation of China (82372037 to H-C Sun and 82503971 to B. Xu), the Special Research Fund for Liver Cancer Diagnosis and Treatment from the China Anti-Cancer Association (H2020-008 to H.-C. Sun), the Clinical Research Special Fund of Zhongshan Hospital, Fudan University (2020ZSLC71 to H.-C. Sun), the 2022 Personalized Medical Incubator Project, through the fund for Precision Medicine Research and Industry Development in SIMQ (EH20 to H.-C. Sun), and Outstanding Resident Clinical Postdoctoral Program of Zhongshan Hospital Affiliated to Fudan University (2025ZYYS-049 to B. Xu).
Data Availability Statement
The data supporting the findings of this study are not publicly available due to patient privacy concerns, but they are available from the corresponding author upon reasonable request.
References
- 1. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209–49. [DOI] [PubMed] [Google Scholar]
- 2. Younossi ZM, Wong G, Anstee QM, Henry L. The global burden of liver disease. Clin Gastroenterol Hepatol. 2023;21(8):1978–91. [DOI] [PubMed] [Google Scholar]
- 3. Zeng H, Cao M, Xia C, Wang D, Chen K, Zhu Z, et al. Performance and effectiveness of hepatocellular carcinoma screening in individuals with HBsAg seropositivity in China: a multicenter prospective study. Nat Cancer. 2023;4(9):1382–94. [DOI] [PubMed] [Google Scholar]
- 4. Peng Z, Fan W, Zhu B, Wang G, Sun J, Xiao C, et al. Lenvatinib combined with transarterial chemoembolization as first-line treatment for advanced hepatocellular carcinoma: a phase III, randomized clinical trial (LAUNCH). J Clin Oncol. 2023;41(1):117–27. [DOI] [PubMed] [Google Scholar]
- 5. He M, Li Q, Zou R, Shen J, Fang W, Tan G, et al. Sorafenib plus hepatic arterial infusion of Oxaliplatin, fluorouracil, and Leucovorin vs sorafenib alone for hepatocellular carcinoma with portal vein invasion: a randomized clinical trial. JAMA Oncol. 2019;5(7):953–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Kudo M, Ren Z, Guo Y, Han G, Lin H, Zheng J, et al. Transarterial chemoembolisation combined with lenvatinib plus pembrolizumab versus dual placebo for unresectable, non-metastatic hepatocellular carcinoma (LEAP-012): a multicentre, randomised, double-blind, phase 3 study. Lancet. 2025;405(10474):203–15. [DOI] [PubMed] [Google Scholar]
- 7. Sangro B, Kudo M, Erinjeri JP, Qin S, Ren Z, Chan SL, et al. Durvalumab with or without bevacizumab with transarterial chemoembolisation in hepatocellular carcinoma (EMERALD-1): a multiregional, randomised, double-blind, placebo-controlled, phase 3 study. Lancet. 2025;405(10474):216–32. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Kudo M, Ueshima K, Ikeda M, Torimura T, Tanabe N, Aikata H, et al. Randomised, multicentre prospective trial of transarterial chemoembolisation (TACE) plus sorafenib as compared with TACE alone in patients with hepatocellular carcinoma: TACTICS trial. Gut. 2020;69(8):1492–501. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Fan W, Zhu B, Chen S, Wu Y, Zhao X, Qiao L, et al. Survival in patients with recurrent intermediate-stage hepatocellular carcinoma: Sorafenib plus TACE vs TACE alone randomized clinical trial. JAMA Oncol. 2024;10(8):1047–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Aoki T, Nishida N, Ueshima K, Morita M, Chishina H, Takita M, et al. Higher enhancement intrahepatic nodules on the hepatobiliary phase of Gd-EOB-DTPA-Enhanced MRI as a poor responsive marker of Anti-PD-1/PD-L1 monotherapy for unresectable hepatocellular carcinoma. Liver Cancer. 2021;10(6):615–28. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Zhang K, Zhang L, Li WC, Xie SS, Cui YZ, Lin LY, et al. Radiomics nomogram for the prediction of microvascular invasion of HCC and patients' benefit from postoperative adjuvant TACE: a multi-center study. Eur Radiol. 2023;33(12):8936–47. [DOI] [PubMed] [Google Scholar]
- 12. Hua Y, Sun Z, Xiao Y, Li H, Ma X, Luo X, et al. Pretreatment CT-based machine learning radiomics model predicts response in unresectable hepatocellular carcinoma treated with lenvatinib plus PD-1 inhibitors and interventional therapy. J Immunother Cancer. 2024;12(7):e008953. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Gatenby RA, Grove O, Gillies RJ. Quantitative imaging in cancer evolution and ecology. Radiology. 2013;269(1):8–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Calderaro J, Seraphin TP, Luedde T, Simon TG. Artificial intelligence for the prevention and clinical management of hepatocellular carcinoma. J Hepatol. 2022;76(6):1348–61. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Zhang Z, Luo T, Yan M, Shen H, Tao K, Zeng J, et al. Voxel-level radiomics and deep learning for predicting pathologic complete response in esophageal squamous cell carcinoma after neoadjuvant immunotherapy and chemotherapy. J Immunother Cancer. 2025;13(3):e011149. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Lan T, Kuang S, Liang P, Ning C, Li Q, Wang L, et al. MRI-based deep learning and radiomics for prediction of occult cervical lymph node metastasis and prognosis in early-stage oral and oropharyngeal squamous cell carcinoma: a diagnostic study. Int J Surg. 2024;110(8):4648–59. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Ye G, Wu G, Li K, Zhang C, Zhuang Y, Liu H, et al. Development and validation of a deep learning radiomics model to predict high-risk pathologic pulmonary nodules using preoperative computed tomography. Acad Radiol. 2024;31(4):1686–97. [DOI] [PubMed] [Google Scholar]
- 18. Lin Z, Wang W, Yan Y, Ma Z, Xiao Z, Mao K. A deep learning-based clinical-radiomics model predicting the treatment response of immune checkpoint inhibitors (ICIs)-based conversion therapy in potentially convertible hepatocelluar carcinoma patients: a tumor marker prognostic study. Int J Surg. 2025;111(5):3342–55. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Shi S, Lin C, Zhou J, Wei L, Chen M, Zhang J, et al. Development and validation of a deep learning radiomics model with clinical-radiological characteristics for the identification of occult peritoneal metastases in patients with pancreatic ductal adenocarcinoma. Int J Surg. 2024;110(5):2669–78. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Zhou J, Sun H, Wang Z, Cong W, Wang J, Zeng M, et al. Guidelines for the diagnosis and treatment of hepatocellular carcinoma (2019 edition). Liver Cancer. 2020;9(6):682–720. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Bruix J, Sherman M; American Association for the Study of Liver Diseases . Management of hepatocellular carcinoma: an update. Hepatology. 2011;53(3):1020–2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Llovet JM, Lencioni R. mRECIST for HCC: performance and novel refinements. J Hepatol. 2020;72(2):288–306. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. [DOI] [PubMed] [Google Scholar]
- 24. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Peng H, Long F, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell. 2005;27(8):1226–38. [DOI] [PubMed] [Google Scholar]
- 26. Lim M, Hastie T. Learning interactions via hierarchical group-lasso regularization. J Comput Graph Stat. 2015;24(3):627–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Wu J, Xia Y, Wang X, Wei Y, Liu A, Innanje A, et al. uRP: an integrated research platform for one-stop analysis of medical images. Front Radiol. 2023;3:1153784. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: 2016 IEEE conference on computer vision and pattern recognition (CVPR). IEEE Computer Society; 2016; p. 770–8. [Google Scholar]
- 29. Kudo M, Guo Y, Hua Y, Zhao M, Xing W, Zhang Y, et al. TALENTACE: a phase III, open-label, randomized study of on-demand transarterial chemoembolization combined with atezolizumab + bevacizumab or on-demand transarterial chemoembolization alone in patients with untreated hepatocellular carcinoma. J Clin Oncol. 2022;40(4_Suppl l):TPS487. [Google Scholar]
- 30. Scheiner B, Pomej K, Kirstein MM, Hucke F, Finkelmeier F, Waidmann O, et al. Prognosis of patients with hepatocellular carcinoma treated with immunotherapy - development and validation of the CRAFITY score. J Hepatol. 2022;76(2):353–63. [DOI] [PubMed] [Google Scholar]
- 31. Zhu AX, Abbas AR, de Galarreta MR, Guan Y, Lu S, Koeppen H, et al. Molecular correlates of clinical response and resistance to atezolizumab in combination with bevacizumab in advanced hepatocellular carcinoma. Nat Med. 2022;28(8):1599–611. [DOI] [PubMed] [Google Scholar]
- 32. Guo X, Nie H, Zhang W, Li J, Ge J, Xie B, et al. Contrasting cytotoxic and regulatory T cell responses underlying distinct clinical outcomes to anti-PD-1 plus lenvatinib therapy in cancer. Cancer Cell. 2025;43(2):248–68.e9. [DOI] [PubMed] [Google Scholar]
- 33. Eun JW, Yoon JH, Ahn HR, Kim S, Kim YB, Lim SB, et al. Cancer-associated fibroblast-derived secreted phosphoprotein 1 contributes to resistance of hepatocellular carcinoma to sorafenib and lenvatinib. Cancer Commun. 2023;43(4):455–79. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Vithayathil M, Koku D, Campani C, Nault JC, Sutter O, Carrie NG, et al. Machine learning based radiomic models outperform clinical biomarkers in predicting outcomes after immunotherapy for hepatocellular carcinoma. J Hepatol. 2025;83(4):959–70. [DOI] [PubMed] [Google Scholar]
- 35. Bo Z, Chen B, Zhao Z, He Q, Mao Y, Yang Y, et al. Prediction of response to lenvatinib monotherapy for unresectable hepatocellular carcinoma by machine learning radiomics: a multicenter cohort study. Clin Cancer Res. 2023;29(9):1730–40. [DOI] [PubMed] [Google Scholar]
- 36. Minami Y, Nishida N, Kudo M. Imaging diagnosis of various hepatocellular carcinoma subtypes and its hypervascular mimics: differential diagnosis based on conventional interpretation and artificial intelligence. Liver Cancer. 2023;12(2):103–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Nam D, Chapiro J, Paradis V, Seraphin TP, Kather JN. Artificial intelligence in liver diseases: improving diagnostics, prognostics and response prediction. JHEP Rep. 2022;4(4):100443. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Chen M, Kong C, Qiao E, Chen Y, Chen W, Jiang X, et al. Multi-algorithms analysis for pre-treatment prediction of response to transarterial chemoembolization in hepatocellular carcinoma on multiphase MRI. Insights Imaging. 2023;14(1):38. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Yan M, Zhang X, Zhang B, Geng Z, Xie C, Yang W, et al. Deep learning nomogram based on Gd-EOB-DTPA MRI for predicting early recurrence in hepatocellular carcinoma after hepatectomy. Eur Radiol. 2023;33(7):4949–61. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Liu B, Zeng Q, Huang J, Zhang J, Zheng Z, Liao Y, et al. IVIM using convolutional neural networks predicts microvascular invasion in HCC. Eur Radiol. 2022;32(10):7185–95. [DOI] [PubMed] [Google Scholar]
- 41. Kim S, Shin J, Kim DY, Choi GH, Kim MJ, Choi JY. Radiomics on gadoxetic acid-enhanced magnetic resonance imaging for prediction of postoperative early and late recurrence of single hepatocellular carcinoma. Clin Cancer Res. 2019;25(13):3847–55. [DOI] [PubMed] [Google Scholar]
- 42. Kong C, Zhao Z, Chen W, Lv X, Shu G, Ye M, et al. Prediction of tumor response via a pretreatment MRI radiomics-based nomogram in HCC treated with TACE. Eur Radiol. 2021;31(10):7500–11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43. Miltiadous O, Sia D, Hoshida Y, Fiel MI, Harrington AN, Thung SN, et al. Progenitor cell markers predict outcome of patients with hepatocellular carcinoma beyond milan criteria undergoing liver transplantation. J Hepatol. 2015;63(6):1368–77. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44. Yang L, Gu D, Wei J, Yang C, Rao S, Wang W, et al. A radiomics nomogram for preoperative prediction of microvascular invasion in hepatocellular carcinoma. Liver Cancer. 2019;8(5):373–86. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45. Wu M, Tan H, Gao F, Hai J, Ning P, Chen J, et al. Predicting the grade of hepatocellular carcinoma based on non-contrast-enhanced MRI radiomics signature. Eur Radiol. 2019;29(6):2802–11. [DOI] [PubMed] [Google Scholar]
- 46. Xu B, Dong SY, Bai XL, Song TQ, Zhang BH, Zhou LD, et al. Tumor radiomic features on pretreatment MRI to predict response to Lenvatinib plus an Anti-PD-1 antibody in advanced hepatocellular carcinoma: a multicenter study. Liver Cancer. 2023;12(3):262–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47. Liao NQ, Deng ZJ, Wei W, Lu JH, Li MJ, Ma L, et al. Deep learning of pretreatment multiphase CT images for predicting response to lenvatinib and immune checkpoint inhibitors in unresectable hepatocellular carcinoma. Comput Struct Biotechnol J. 2024;24:247–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data supporting the findings of this study are not publicly available due to patient privacy concerns, but they are available from the corresponding author upon reasonable request.