Published online Oct 15, 2026. doi: 10.4251/wjgo.122464
Revised: June 11, 2026
Accepted: July 30, 2026
Published online: October 15, 2026
Processing time: 172 Days and 19.4 Hours
Early recurrence (ER) remains a major challenge after ablation therapy for hepatocellular carcinoma (HCC) and is closely associated with poor prognosis. Peritu
To determine the best-performing peritumoral region for predicting ER after HCC ablation using DL features.
This retrospective cohort study enrolled 296 ablation-treated HCC patients from two centers. The patients were divided into a primary cohort (n = 222) and an external cohort (n = 74). The 3D ResNet-18 was used to extract DL features from contrast-enhanced computed tomography images of the tumor and the 3-mm, 5-mm, 10-mm, and 20-mm peritumoral regions. ER-related features were selected with least absolute shrinkage and selection operator regression, recursive feature elimination, and importance ranking. Six machine learning algorithms were used to construct predictive models to determine the region whose features resulted in optimal performance. A combined model incorporating the optimal peritumoral features with clinical imaging features (postoperative neutrophil count, intratumoral necrosis) was also developed. The models were evaluated with receiver operating characteristic curve analysis [including calculation of the area under the curve (AUC)], calibration/decision curve analysis, and Kaplan-Meier survival analysis.
The 10-mm peritumoral model outperformed both the tumor model and the other peritumoral models, with AUC values of 0.818 (95%CI: 0.802-0.833), 0.804 (95%CI: 0.779-0.829) in the training and internal validation sets, respectively. Incorporation of the clinical features into a combined model with the peritumoral features improved the predictive accuracy further, with AUCs of 0.855 (95%CI: 0.841-0.869), 0.852 (95%CI: 0.832-0.871) and 0.841 (95%CI: 0.811-0.871) in the training, internal and external validation sets, respectively. Calibration curve and decision curve analysis also indicated favorable performance with the combined model. Kaplan-Meier analysis demonstrated that the combined model effectively stratified patients in terms of progression-free survival and overall survival probabilities.
DL features extracted from the 10-mm peritumoral region showed the best numerical predictive performance for ER in HCC patients after ablation. The combined model demonstrated favorable predictive performance and effectively stratified patients by survival risk and thus could aid in developing personalized treatment strategies.
Core Tip: This study developed a deep learning-enhanced peritumoral radiomics model to predict early recurrence after hepatocellular carcinoma ablation. Among the evaluated peritumoral regions, the 10-mm peritumoral model showed the best numerical performance. The combined model integrating 10-mm peritumoral deep learning features and clinical imaging features enabled individualized recurrence risk stratification and may aid in developing personalized treatment strategies.
- Citation: Li YH, Yao L, Qian GX, Lei XD, Tang ZQ, Guo BY, Du R, Zhu Y, Jia WD. Deep learning-enhanced peritumoral radiomics predicts early recurrence after hepatocellular carcinoma ablation: A two-center study. World J Gastrointest Oncol 2026; 18(10): 122464
- URL: https://www.wjgnet.com/1948-5204/full/v18/i10/122464.htm
- DOI: https://dx.doi.org/10.4251/wjgo.122464
Hepatocellular carcinoma (HCC) is the sixth most prevalent cancer worldwide and the third leading cause of cancer-related mortality[1,2]. Furthermore, it is the most predominant pathological subtype of primary liver cancer, accounting for 75%-85% of all cases. For patients diagnosed with early-stage HCC, current clinical guidelines recommend curative interventions, such as surgical resection, liver transplantation, and local ablation therapy. However, only approximately 30% of such patients are eligible for liver resection, and liver transplantation remains limited by organ shortages, therefore benefiting a limited proportion of patients[3-5]. Recent evidence further indicates that complex hepatectomy involving major hepatic venous structures may increase operative difficulty and perioperative burden, highlighting the limitations of surgical treatment in selected patients[6]. In contrast, ablation therapy has gained prominence as an effective treatment modality for HCC, with comparable efficacy to surgical resection, fewer complications, and reduced costs. Despite these advantages, however, the recurrence rates post-HCC ablation remain high, ranging from 21.9% to 70.2%[7,8]. Recurrence, a significant prognostic factor in patients with HCC, is generally classified into early recurrence (ER) and late recurrence, between which ER is frequently associated with poorer survival outcomes than late recurrence[9].
While most current research on recurrence has focused primarily on its likelihood following surgical resection, evidence suggests that the presence of nonviable tumor cells postablation may lead to the release of tumor antigens, potentially functioning as autologous vaccines that could mitigate recurrence[10]. This hypothesis suggests that recurrence patterns following ablation may differ from those observed after liver resection. Consequently, a deeper understanding of ER following HCC ablation is crucial for optimizing treatment protocols and follow-up strategies.
Radiomics, an imaging technique that involves the extraction of quantitative features from medical imaging data, has emerged as an important tool for predicting tumor behavior[11]. The advent of deep learning (DL), particularly in the form of convolutional neural networks, has notably improved the predictive capacity of radiomic feature-based models[12,13]. Pan and Yang[14] demonstrated the effectiveness of radiomic models that leverage DL features in diagnosing focal liver lesions. Liu et al[15] developed a DL-based radiomic model that accurately predicted progression-free survival (PFS) in early-stage HCC patients undergoing radiofrequency ablation or liver resection, with C indices of 0.754 and 0.726 in the training and validation cohorts, respectively. Similarly, Guo et al[16] developed a predictive model by integrating DL features with the clinical HCC risk score, which provided a reliable early warning of HCC in patients with liver cirrhosis. The model achieved area under the curve (AUC) values of 0.929, 0.902, and 0.918 in the training, internal validation, and external validation sets, respectively.
Tumor biology is shaped by both intratumoral and peritumoral characteristics, with the latter playing a pivotal role in tumor progression and patient outcomes[17,18]. Imaging features within a 5-mm peritumoral region, for instance, have been shown to be a promising noninvasive biomarker for preoperatively predicting microvascular invasion in patients diagnosed with a solitary HCC ≤ 5 cm.[19,20]. Kang et al[21] demonstrated that the features extracted from contrast-enhanced computed tomography (CECT) images of a 3-mm peritumoral region were the most reliable predictors of ER following liver resection (AUC = 0.807). Additionally, Liu et al[22] reported that a 10-mm peritumoral region feature-based model performed best in predicting the patients’ HCC pathological subtype. Despite the valuable insights gleaned from DL-based radiomic models, systematic investigations into the optimal peritumoral range for predicting ER following HCC ablation remain limited.
This retrospective cohort study was approved by the Ethics Committee of the First Affiliated Hospital of University of Science and Technology of China and was conducted in accordance with the ethical guidelines of the 1975 Helsinki Declaration. The study was exempt from the requirement for patient informed consent (approval No. 2021-RE-043). A total of 296 patients diagnosed with HCC who underwent ablation therapy between April 2008 and August 2021 at the First Affiliated Hospital of University of Science and Technology of China and Taizhou Hospital of Zhejiang Province Affiliated to Wenzhou Medical University from December 2018 to December 2022 were included. The selection process is illustrated in Figure 1.
Inclusion criteria: (1) Ablation therapy (radiofrequency or microwave); (2) A clinical diagnosis of HCC made according to noninvasive, imaging characteristic-based criteria defined by the American Association for the Study of Liver Diseases[23]; (3) Ineligibility for or unwillingness to undergo liver resection or transplantation; and (4) A Child-Pugh classification of grade A.
Exclusion criteria: (1) The presence of vascular, bile duct, or adjacent organ invasion; distant metastasis; intrahepatic vascular tumor thrombus; or concurrent other malignancies; (2) A single tumor > 5 cm in size or more than three tumors; (3) A previous diagnosis or treatment of HCC; (4) A lack of CECT images or no CECT imaging within one month prior to ablation; (5) Failed or incomplete ablation according to the postprocedure evaluation; and (6) A follow-up duration of < 2 years. The detailed inclusion and exclusion process is shown in Figure 2.
Ablation was performed by a team of experienced physicians under ultrasound guidance. Prior to the procedure, the patients underwent a comprehensive evaluation involving determination of tumor size and location to select the ablation technique (radiofrequency or microwave). The optimal puncture site was determined under ultrasound guidance to ensure precise needle placement at the site of the target lesion. During the procedure, the ablation zone, visible as a hyperechoic region on ultrasound, was continuously monitored to ensure complete tumor coverage. After ablation, ultrasound imaging was again used to assess the adequacy of the ablation. Complete ablation was defined as the creation of a margin of at least 5 mm of liver tissue surrounding the tumor.
The CECT images analyzed in this study were retrieved in DICOM format from the hospital’s picture archiving and communication system. Details of the CT equipment, acquisition protocols, and standardized image preprocessing procedures are provided in Supplementary material. A radiologist with 5 years of experience in abdominal imaging manually delineated the regions of interests (ROIs) in the arterial (A), portal-venous (P), and delayed (D) phases using ITK-SNAP (v3.6.0) under the supervision of a senior radiologist with 10 years of experience. Neither physician was aware of the patients’ clinical details during segmentation. To evaluate the predictive capability of features from differently sized peritumoral regions, the tumor ROIs were manually delineated, after which an automatic expansion algorithm in Python was used to generate the 3-mm, 5-mm, 10-mm, and 20-mm peritumoral regions (Supplementary Figure 1). Nonli
A ResNet-18 model initially trained on the ImageNet dataset was employed for transfer learning[13]. The Adam optimizer was used to fine-tune the network parameters. To adapt the model for deep feature extraction, the final fully connected layer was removed, and a 512-dimensional feature vector was extracted from each input region of interest. The detailed implementation of the modified 3D ResNet-18 model is provided in Supplementary material.
Clinical imaging features were selected using both univariable and multivariable methods. DL features underwent a three-step selection process comprising least absolute shrinkage and selection operator regression, recursive feature elimination, and importance ranking.
Least absolute shrinkage and selection operator regression was first applied for preliminary dimensionality reduction. Testing a wide range of α values (10-7-10) revealed that larger α values removed nearly all features; thus, α = 0.001 was chosen to avoid over-penalization while retaining informative variables for subsequent steps. Recursive feature elimination with a decision tree classifier was then used to refine the feature set, followed by importance ranking, where the top ten features were retained to enhance robustness. Ultimately, five feature sets were obtained, one for each (peri)tumoral region. To assess stability, the procedure was repeated within five-fold cross-validation.
Six machine learning algorithms were employed to construct predictive models on the basis of the DL radiomic features: (1) Support vector machine; (2) Logistic regression; (3) Random forest; (4) K-nearest neighbors; (5) Light gradient boosting machine; and (6) Extreme gradient boosting. The best-performing algorithm was identified on the basis of the per
Following discharge, the patients were scheduled for routine follow-up appointments. The follow-up evaluations consisted of verification of complete ablation one-month postprocedure, as well as blood tests (alpha fetoprotein, liver function) and abdominal ultrasounds every 3-6 months. If recurrence was suspected, CECT, magnetic resonance imaging, or contrast-enhanced ultrasound was performed. ER was defined as the emergence of new intrahepatic lesions or metastases within two years postablation, as indicated by typical HCC imaging features or confirmed through histopathological examination. The follow-up period ended on March 31, 2025.
Variables with more than 20% missing values were excluded from the analysis. For variables with less than 20% missing values, multiple imputation techniques were used to impute these missing values. All the statistical analyses were conducted in Python (version 3.6; Python Software Foundation, Wilmington, DE, United States) and R (version 4.2.2; R Foundation for Statistical Computing, Vienna, Austria). Categorical variables were compared between groups Pearson’s χ2 test or Fisher’s exact test, whereas continuous variables were compared between groups with Student’s t test or Mann-Whitney U test. Normally distributed data are expressed as the means ± SD, and nonnormally distributed data are expressed as the medians (interquartile ranges). Model performance was evaluated with the AUC. Calibration and clinical applicability were assessed using calibration curves and decision curve analysis (DCA), respectively. Survival was analyzed with using Kaplan-Meier curve analysis, and comparisons between the groups were made with the log-rank test. A two-sided P value < 0.05 was considered to indicate statistical significance.
A total of 296 patients were included in this study, among whom 40.0% in the training cohort, 38.8% in the internal validation cohort and 33.8% in the external validation cohort experienced ER. No significant differences were observed in the baseline characteristics between the three groups. The detailed data are presented in Table 1.
| Item | Training (n = 155) | Internal validation (n = 67) | External validation (n = 74) | P value |
| Tumor size (cm) | 2.91 (1.64) | 2.68 (0.98) | 2.70 (1.18) | 0.408 |
| Age (years) | 57.9 (11.1) | 58.8 (12.4) | 58.6 (12.0) | 0.853 |
| Sex | 0.833 | |||
| Female | 30 (19.4) | 13 (19.4) | 12 (16.2) | |
| Male | 125 (80.6) | 54 (80.6) | 62 (83.8) | |
| HBV | 0.435 | |||
| Positive | 53 (34.2) | 29 (43.3) | 27 (36.5) | |
| Negative | 102 (65.8) | 38 (56.7) | 47 (63.5) | |
| Cirrhosis | 0.261 | |||
| Positive | 28 (18.1) | 17 (25.4) | 11 (14.9) | |
| Negative | 127 (81.9) | 50 (74.6) | 63 (85.1) | |
| N | 0.949 | |||
| < 1.8 | 64 (41.3) | 24 (35.8) | 28 (37.8) | |
| 1.8-6.3 | 89 (57.4) | 42 (62.7) | 45 (60.8) | |
| > 6.3 | 2 (1.29) | 1 (1.49) | 1 (1.35) | |
| L | 0.583 | |||
| ≥ 1.1 | 67 (43.2) | 31 (46.3) | 28 (37.8) | |
| < 1.1 | 88 (56.8) | 36 (53.7) | 46 (62.2) | |
| ALT | 0.085 | |||
| ≤ 50 | 115 (74.2) | 47 (70.1) | 63 (85.1) | |
| > 50 | 40 (25.8) | 20 (29.9) | 11 (14.9) | |
| AST | 0.347 | |||
| > 40 | 68 (43.9) | 27 (40.3) | 25 (33.8) | |
| ≤ 40 | 87 (56.1) | 40 (59.7) | 49 (66.2) | |
| GGT | 0.805 | |||
| ≤ 60 | 85 (54.8) | 38 (56.7) | 38 (51.4) | |
| > 60 | 70 (45.2) | 29 (43.3) | 36 (48.6) | |
| GLB | 0.444 | |||
| ≤ 35 | 122 (78.7) | 52 (77.6) | 63 (85.1) | |
| > 35 | 33 (21.3) | 15 (22.4) | 11 (14.9) | |
| HbsAg | 0.757 | |||
| Positive | 28 (18.1) | 13 (19.4) | 11 (14.9) | |
| Negative | 127 (81.9) | 54 (80.6) | 63 (85.1) | |
| AFP | 0.167 | |||
| < 40 | 39 (25.2) | 20 (29.9) | 15 (20.3) | |
| 40-400 | 92 (59.4) | 30 (44.8) | 47 (63.5) | |
| > 400 | 24 (15.5) | 17 (25.4) | 12 (16.2) | |
| Np | 0.933 | |||
| < 1.8 | 31 (20.0) | 10 (14.9) | 14 (18.9) | |
| 1.8-6.3 | 70 (45.2) | 33 (49.3) | 34 (45.9) | |
| > 6.3 | 54 (34.8) | 24 (35.8) | 26 (35.1) | |
| Lp | 0.968 | |||
| ≥ 1.1 | 101 (65.2) | 43 (64.2) | 49 (66.2) | |
| < 1.1 | 54 (34.8) | 24 (35.8) | 25 (33.8) | |
| ALTp | 0.504 | |||
| ≤ 50 | 33 (21.3) | 13 (19.4) | 20 (27.0) | |
| > 50 | 122 (78.7) | 54 (80.6) | 54 (73.0) | |
| ASTp | 0.483 | |||
| > 40 | 136 (87.7) | 62 (92.5) | 64 (86.5) | |
| ≤ 40 | 19 (12.3) | 5 (7.46) | 10 (13.5) | |
| GGTp | 0.965 | |||
| ≤ 60 | 92 (59.4) | 41 (61.2) | 44 (59.5) | |
| > 60 | 63 (40.6) | 26 (38.8) | 30 (40.5) | |
| GLBp | 0.972 | |||
| ≤ 35 | 115 (74.2) | 50 (74.6) | 54 (73.0) | |
| > 35 | 40 (25.8) | 17 (25.4) | 20 (27.0) | |
| Capsule appearance | 0.136 | |||
| Positive | 97 (62.6) | 39 (58.2) | 36 (48.6) | |
| Negative | 58 (37.4) | 28 (41.8) | 38 (51.4) | |
| Intratumor vascularity | 0.811 | |||
| Positive | 60 (38.7) | 28 (41.8) | 27 (36.5) | |
| Negative | 95 (61.3) | 39 (58.2) | 47 (63.5) | |
| Tumor growth patter | 0.612 | |||
| Positive | 113 (72.9) | 53 (79.1) | 56 (75.7) | |
| Negative | 42 (27.1) | 14 (20.9) | 18 (24.3) | |
| Intratumor necrosis | 0.781 | |||
| Positive | 92 (59.4) | 39 (58.2) | 47 (63.5) | |
| Negative | 63 (40.6) | 28 (41.8) | 27 (36.5) | |
| PACE | 0.607 | |||
| Positive | 123 (79.4) | 54 (80.6) | 55 (74.3) | |
| Negative | 32 (20.6) | 13 (19.4) | 19 (25.7) | |
| Tumor margin | 0.648 | |||
| Positive | 100 (64.5) | 39 (58.2) | 45 (60.8) | |
| Negative | 55 (35.5) | 28 (41.8) | 29 (39.2) | |
| Ablation method | 0398 | |||
| RFA | 44 (28.4) | 21 (31.3) | 16 (21.6) | |
| MWA | 111 (71.6) | 46 (68.7) | 58 (7.4) | |
| PFS | 0.659 | |||
| Absent | 93 (60.0) | 41 (61.2) | 49 (66.2) | |
| Present | 62 (40.0) | 26 (38.8) | 25 (33.8) |
To identify clinical imaging features predictive of ER following HCC ablation, both univariable and multivariable methods were employed for feature selection. Initially, a broad set of clinical variables – including alpha fetoprotein, liver function indices, cirrhosis status, and tumor size (Table 1) – were evaluated. However, these variables did not retain statistical significance in the univariable analysis and were therefore excluded from the final model. Ultimately, the postoperative neutrophil (Np) count and intratumoral necrosis were found to be significantly associated with ER, as detailed in Table 2 and Supplementary Table 1.
| Name | Univariable | Multivariable | ||||
| OR | 95%CI | P value | OR | 95%CI | P value | |
| Np | ||||||
| < 1.8 | - | - | - | - | - | - |
| 1.8-6.3 | 0.70 | 0.30-1.64 | 0.416 | 0.72 | 0.29-1.77 | 0.476 |
| > 6.3 | 0.39 | 0.16-0.99 | 0.046 | 0.29 | 0.10-0.80 | 0.017 |
| IV | ||||||
| Positive | - | - | - | - | - | - |
| Negative | 2.28 | 1.14-4.54 | 0.02 | 1.67 | 0.78-3.57 | 0.19 |
| IN | ||||||
| Positive | - | - | - | - | - | - |
| Negative | 2.13 | 1.10-4.12 | 0.024 | 2.55 | 1.17-5.57 | 0.018 |
A clinical imaging model incorporating Np and intratumoral necrosis was subsequently developed. The model achieved an AUC of 0.605 (95%CI: 0.589-0.621) in the training set and 0.615 (95%CI: 0.592-0.639) in the validation set. These results suggest that while the model based on clinical imaging features possesses a certain amount of predictive power, further refinement could improve its accuracy.
Following a three-step feature selection process, 10 features were extracted from the tumor and each of the peritumoral regions (3 mm, 5 mm, 10 mm, and 20 mm). To identify the optimal ROI for prediction, we first constructed models on the basis of the 10 features from each region and compared their performance. The model utilizing features from the 10-mm peritumoral region demonstrated the best predictive performance, with an AUC of 0.818 (95%CI: 0.802-0.833) in the training set, 0.804 (95%CI: 0.779-0.829) in the internal validation set (Figure 3A and B). To assess stability, the procedure was repeated within five-fold cross-validation, Of the final ten features, six were consistently selected in all folds, three in four folds, and one in three folds, confirming their reproducibility (Supplementary Table 2). The detailed performance of all models is provided in the Supplementary Table 3.
To improve feature-level interpretability, SHapley Additive exPlanations (SHAP) analysis was performed for the 10 selected DL features derived from the 10-mm peritumoral region. The SHAP plot shown in Figure 4 illustrates the relative contribution of each feature to ER prediction.
To assess whether the integration of DL and clinical imaging features could improve model performance, a combined model was developed. The performance of this combined model was then compared with that of the original clinical imaging model and the 10-mm peritumoral model. In the training set, the AUCs of the three models were 0.605, 0.818, and 0.855; in the internal validation set, the AUCs were 0.615, 0.804 and 0.852; in the external validation set, the AUCs were 0.606, 0.805 and 0.841 (Figure 5). These results indicate that the combined model exhibited the highest predictive performance among the three validation models.
Furthermore, the DeLong test was conducted to statistically compare the predictive efficacy of the clinical imaging model, the 10-mm peritumoral model, and the combined model. In both the training and validation sets, the model based on the 10-mm peritumoral features demonstrated significantly greater performance in terms of the AUC than the clinical imaging model (P < 0.05 for all sets). The combined model, which incorporated both the 10-mm peritumoral features and the clinical imaging features, outperformed the 10-mm peritumoral model in terms of the AUC; however, the difference was not significant (P > 0.05 for the training, internal and external validation sets). The detailed results of the DeLong test for comparing the models in the validation sets are provided in Table 3 and Supplementary Table 4.
| Model | Training | Internal validation | External validation | ||||
| Z value | P value | Z value | P value | Z value | P value | ||
| Cli | 10 mm | -3.916 | < 0.0001 | -2.310 | 0.0208 | -2.212 | 0.0230 |
| Cli | Combined | -4.948 | < 0.0001 | -3.498 | 0.0005 | -3.035 | 0.0024 |
| 10 mm | Combined | -1.086 | 0.2773 | -1.063 | 0.2878 | -0.798 | 0.4249 |
In addition, to further assess the robustness of the combined model, we performed bootstrapping with 1000 iterations. The results of these robustness checks are presented in Supplementary Figure 2.
To assess the calibration and clinical applicability of the combined model, calibration curves were plotted, and DCA was performed, respectively (Figure 6). The results indicated that the combined model demonstrated good calibration and offers good net clinical benefits.
The maximum Youden index (0.573) derived from the combined model in the training set was used as the optimal cutoff value. This cutoff was applied to the patients in both the training, internal validation and external cohorts to stratify them into low-risk and high-risk groups. Kaplan-Meier survival curves were generated to analyze the 2-year PFS and 5-year overall survival (OS) rates of both groups (Figure 7). The results demonstrated that, in both the training set (Figure 7A and D), the internal validation set (Figure 7B and E) and the external validation set (Figure 7C and F), the combined model effectively stratified patients into low-risk and high-risk groups on the basis of these survival rates (P < 0.05).
In this study, DL features extracted from three-phase CECT images of both tumor and peritumoral regions were used to construct radiomic models for predicting ER after HCC ablation. Furthermore, a combined model integrating these DL features with clinical imaging features significantly improved the predictive performance, achieving an AUC of 0.855 (95%CI: 0.841-0.869) in the training set, 0.852 (95%CI: 0.832-0.871) in the internal validation set and 0.841 (95%CI: 0.811-0.871) in the external validation set, surpassing the models based on only the clinical imaging features or the 10-mm peritumoral features. Calibration curves and DCA demonstrated that the combined model exhibited favorable calibration and strong clinical applicability, respectively. Additionally, the model effectively stratified patients into high-risk and low-risk groups in terms of the 2-year PFS and 5-year OS rates.
The proportion of patients who experienced ER in this study, 38.18%, is consistent with that previously reported in the literature (range 28.46%-57.66%)[24,25]. We specifically focused on patients with newly diagnosed early-stage HCC who underwent radical ablation, ensuring homogeneity in the data and accuracy in imaging feature extraction and establishing a robust foundation for developing a reliable recurrence prediction model. The use of the 3D ResNet-18 architecture for extracting 3D imaging features from both the tumor and peritumoral regions was a strength of this study. The simplified network design of ResNet-18 results in minimal computational complexity while maintaining high operational efficiency, making the architecture well suited for further developing models for real-time clinical prediction. Previous research by Chen et al[25] demonstrated that DL models based on ResNet-18 excel at predicting local tumor progression following thermal ablation in liver cancer, with notably higher AUC values than models relying solely on clinical variables (P < 0.001). For this study, traditional machine learning classifiers were employed, as they are more suitable for medium-sized datasets and do not require the high computational resources typically associated with DL approaches. In this regard, our study aligns with that by Wu et al[24], who reported that DL models based on ResNet-18-extracted features outperformed traditional radiomic models, offering an efficient and practical predictive solutions.
Univariable and multivariable analyses identified postoperative Np and intratumoral necrosis as independent predictors of ER after HCC ablation. Neutrophils may promote tumor progression by inducing angiogenesis, suppressing antitumor immunity, and shaping a protumor inflammatory microenvironment[26]. Moreover, neutrophil extracellular traps can enhance the metastatic potential of HCC and have been associated with poorer recurrence-free and OS in primary HCC[27,28]. These findings may partly explain the association between postoperative NP and ER observed in our study. Intratumoral necrosis, which typically manifests as hypodense lesions on CECT imaging, has also been shown to predict tumor aggression. Although not directly linked to HCC invasiveness, necrosis is associated with an inadequate blood supply, abnormal angiogenesis, and a hypoxic microenvironment, all of which can contribute to ER. Wang et al[29] demonstrated that intratumoral necrosis is an independent risk factor for early intrahepatic recurrence, further sup
In radiomics, peritumoral features have garnered significant attention because of their role in tumor development, progression, and response to treatment[30]. Although several studies have investigated the predictive value of tumor and peritumoral features for ER in HCC patients, most have focused on patients who have undergone liver resection, and limited research has addressed the role of these features in predicting ER in patients who have undergone ablation therapy. Although Yuan et al[31] extracted features from CECT images to predict ER after HCC ablation and achieved an impressive C-index of 0.736 in the validation set, it included only tumor features and did not incorporate features from peritumoral regions or develop DL methods, leading to lower predictive performance. Wang et al[32] combined mul
Both the combined model, which integrates DL features from the 10-mm peritumoral region with clinical imaging features, and the 10-mm peritumoral model itself outperformed the clinical model (P = 0.021 and P < 0.001, respectively). However, although it demonstrated a higher AUC and thus superior predictive performance, the difference between the combined model and the 10-mm peritumoral model was not statistically significant (P = 0.288). Calibration curves and DCA confirmed the favorable clinical applicability of the combined model, particularly in terms of risk stratification. Zhao et al[38] also reported that integrating DL features with clinical data improved the prediction of ER after HCC surgery over that with the individual sets of features. The performance improvement observed in our study can likely be attributed to the complementary nature of the two groups of features: While the 10-mm peritumoral radiomic features capture local microenvironmental information, the clinical features reflect the patient's overall health status. This integration of local and systemic information provides a more comprehensive prediction of HCC recurrence, offering new insights into the underlying mechanisms involved in this pathology. Thus, the combined model has significant potential for future research and clinical applications.
The proposed model may support individualized risk stratification after HCC ablation by identifying high-risk patients who may require closer surveillance or individualized management, while avoiding unnecessary interventions in low-risk patients.
Nevertheless, several limitations should be acknowledged. First, because of the retrospective design, no formal prospe
In summary, DL features extracted from images of the 10-mm peritumoral region showed the best numerical predictive performance for ER in HCC patients after ablation. Furthermore, the combined model incorporating 10-mm peritumoral and clinical imaging features showed numerically higher predictive accuracy than the individual models, although its improvement over the 10-mm peritumoral model was not statistically significant. Use of this combined model could potentially guide the development of personalized treatment strategies for HCC patients undergoing ablation therapy.
The authors would like to thank Professor Chao Wei and Ying-Ming Zhao from the Department of Radiology, The First Affiliated Hospital of University of Science and Technology of China, Division of Life Sciences and Medicine, University of Science and Technology of China for their assistance in imaging feature identification and image annotation in writing this article.
| 1. | Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2024;74:229-263. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 16785] [Cited by in RCA: 17649] [Article Influence: 8824.5] [Reference Citation Analysis (37)] |
| 2. | Hepatocellular carcinoma. Nat Rev Dis Primers. 2021;7:7. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 217] [Cited by in RCA: 183] [Article Influence: 36.6] [Reference Citation Analysis (6)] |
| 3. | Zhao W, Li H, Li W, Guo J, Cai L, Duan Y, Hou X, Diao Z, Shao X, Du H, Li W, Li C. Effect of microwave ablation on platelet and coagulation function in patients with BCLC-A hepatocellular carcinoma. J Cancer Res Ther. 2021;17:1275-1280. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 3] [Reference Citation Analysis (0)] |
| 4. | Qi E, Zhang S, Li X, Cheng Z, Han Z, Yu J, Liang P, Yu X. Comparison of percutaneous microwave ablation and surgical resection for hepatocellular carcinoma in the caudate lobe. J Cancer Res Ther. 2022;18:378-383. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 12] [Reference Citation Analysis (0)] |
| 5. | Yu J, Liang P. Status and advancement of microwave ablation in China. Int J Hyperthermia. 2017;33:278-287. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 15] [Cited by in RCA: 25] [Article Influence: 2.5] [Reference Citation Analysis (0)] |
| 6. | Luo WL, Wang QB, Li YK, Liang YB, Li J, Chen XM, Lakang Y, Yang ZS, Zuo JX, Wang W, Li SX, Ke Y. Impact of Middle Hepatic Vein Resection During Hemihepatectomy on Surgical Outcomes and Long-Term Prognosis in Hepatocellular Carcinoma: A Retrospective Study. J Hepatocell Carcinoma. 2025;12:2681-2692. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 46] [Reference Citation Analysis (0)] |
| 7. | Zhao JW, Shu X, Chen XX, Liu JX, Liu MQ, Ye J, Jiang HJ, Wang GS. Prediction of early recurrence of hepatocellular carcinoma after liver transplantation based on computed tomography radiomics nomogram. Hepatobiliary Pancreat Dis Int. 2022;21:543-550. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 1] [Cited by in RCA: 14] [Article Influence: 3.5] [Reference Citation Analysis (5)] |
| 8. | Shen JX, Zhou Q, Chen ZH, Chen QF, Chen SL, Feng ST, Li X, Wu TF, Peng S, Kuang M. Longitudinal radiomics algorithm of posttreatment computed tomography images for early detecting recurrence of hepatocellular carcinoma after resection or ablation. Transl Oncol. 2021;14:100866. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 7] [Cited by in RCA: 15] [Article Influence: 2.5] [Reference Citation Analysis (4)] |
| 9. | Zhang J, Guo G, Li T, Guo C, Han Y, Zhou X. Early versus Late Recurrence of Hepatocellular Carcinoma after Microwave Ablation: Patterns, Treatments, and Post-Recurrence Survival. Oncology. 2024;102:785-793. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 1] [Reference Citation Analysis (0)] |
| 10. | Xie GL, Zhong ZH, Ye TW, Xiao ZQ. Radiofrequency ablation combined with immunotherapy to treat hepatocellular carcinoma: a comprehensive review. BMC Surg. 2025;25:47. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 7] [Reference Citation Analysis (0)] |
| 11. | Ji GW, Zhu FP, Xu Q, Wang K, Wu MY, Tang WW, Li XC, Wang XH. Radiomic Features at Contrast-enhanced CT Predict Recurrence in Early Stage Hepatocellular Carcinoma: A Multi-Institutional Study. Radiology. 2020;294:568-579. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 247] [Cited by in RCA: 232] [Article Influence: 38.7] [Reference Citation Analysis (13)] |
| 12. | Liu Z, Wang S, Dong D, Wei J, Fang C, Zhou X, Sun K, Li L, Li B, Wang M, Tian J. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges. Theranostics. 2019;9:1303-1322. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 841] [Cited by in RCA: 724] [Article Influence: 103.4] [Reference Citation Analysis (7)] |
| 13. | Zhang H, Huo F. Prediction of early recurrence of HCC after hepatectomy by contrast-enhanced ultrasound-based deep learning radiomics. Front Oncol. 2022;12:930458. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 14] [Cited by in RCA: 13] [Article Influence: 3.3] [Reference Citation Analysis (0)] |
| 14. | Pan SJ, Yang Q. A Survey on Transfer Learning. IEEE Trans Knowl Data Eng. 2010;22:1345-1359. [DOI] [Full Text] |
| 15. | Liu F, Liu D, Wang K, Xie X, Su L, Kuang M, Huang G, Peng B, Wang Y, Lin M, Tian J, Xie X. Deep Learning Radiomics Based on Contrast-Enhanced Ultrasound Might Optimize Curative Treatments for Very-Early or Early-Stage Hepatocellular Carcinoma Patients. Liver Cancer. 2020;9:397-413. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 127] [Cited by in RCA: 113] [Article Influence: 18.8] [Reference Citation Analysis (6)] |
| 16. | Guo L, Hao X, Chen L, Qian Y, Wang C, Liu X, Fan X, Jiang G, Zheng D, Gao P, Bai H, Wang C, Yu Y, Dai W, Gao Y, Liang X, Liu J, Sun J, Tian J, Wang H, Hou J, Fan R. Early warning of hepatocellular carcinoma in cirrhotic patients by three-phase CT-based deep learning radiomics model: a retrospective, multicentre, cohort study. EClinicalMedicine. 2024;74:102718. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 5] [Cited by in RCA: 16] [Article Influence: 8.0] [Reference Citation Analysis (0)] |
| 17. | Gerlinger M, Rowan AJ, Horswell S, Math M, Larkin J, Endesfelder D, Gronroos E, Martinez P, Matthews N, Stewart A, Tarpey P, Varela I, Phillimore B, Begum S, McDonald NQ, Butler A, Jones D, Raine K, Latimer C, Santos CR, Nohadani M, Eklund AC, Spencer-Dene B, Clark G, Pickering L, Stamp G, Gore M, Szallasi Z, Downward J, Futreal PA, Swanton C. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. N Engl J Med. 2012;366:883-892. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 6687] [Cited by in RCA: 6138] [Article Influence: 438.4] [Reference Citation Analysis (14)] |
| 18. | Tekpli X, Lien T, Røssevold AH, Nebdal D, Borgen E, Ohnstad HO, Kyte JA, Vallon-Christersson J, Fongaard M, Due EU, Svartdal LG, Sveli MAT, Garred Ø; OSBREAC, Frigessi A, Sahlberg KK, Sørlie T, Russnes HG, Naume B, Kristensen VN. An independent poor-prognosis subtype of breast cancer defined by a distinct tumor immune microenvironment. Nat Commun. 2019;10:5499. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 166] [Cited by in RCA: 149] [Article Influence: 21.3] [Reference Citation Analysis (2)] |
| 19. | Wang F, Cheng M, Du B, Li J, Li L, Huang W, Gao J. Predicting microvascular invasion in small (≤ 5 cm) hepatocellular carcinomas using radiomics-based peritumoral analysis. Insights Imaging. 2024;15:90. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 19] [Reference Citation Analysis (1)] |
| 20. | Zhang L, Cai P, Hou J, Luo M, Li Y, Jiang X. Radiomics Model Based on Gadoxetic Acid Disodium-Enhanced MR Imaging to Predict Hepatocellular Carcinoma Recurrence After Curative Ablation. Cancer Manag Res. 2021;13:2785-2796. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 21] [Cited by in RCA: 20] [Article Influence: 4.0] [Reference Citation Analysis (4)] |
| 21. | Kang W, Cao X, Luo J. Effect of multiple peritumoral regions of interest ranges based on computed tomography radiomics for the prediction of early recurrence of hepatocellular carcinoma after resection. Quant Imaging Med Surg. 2023;13:6668-6682. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 9] [Reference Citation Analysis (12)] |
| 22. | Liu HF, Wang M, Wang Q, Lu Y, Lu YJ, Sheng Y, Xing F, Zhang JL, Yu SN, Xing W. Multiparametric MRI-based intratumoral and peritumoral radiomics for predicting the pathological differentiation of hepatocellular carcinoma. Insights Imaging. 2024;15:97. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 11] [Cited by in RCA: 37] [Article Influence: 18.5] [Reference Citation Analysis (10)] |
| 23. | Heimbach JK, Kulik LM, Finn RS, Sirlin CB, Abecassis MM, Roberts LR, Zhu AX, Murad MH, Marrero JA. AASLD guidelines for the treatment of hepatocellular carcinoma. Hepatology. 2018;67:358-380. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 3374] [Cited by in RCA: 3231] [Article Influence: 403.9] [Reference Citation Analysis (11)] |
| 24. | Wu JP, Ding WZ, Wang YL, Liu S, Zhang XQ, Yang Q, Cai WJ, Yu XL, Liu FY, Kong D, Zhong H, Yu J, Liang P. Radiomics analysis of ultrasound to predict recurrence of hepatocellular carcinoma after microwave ablation. Int J Hyperthermia. 2022;39:595-604. [RCA] [PubMed] [DOI] [Full Text] [Cited by in RCA: 28] [Reference Citation Analysis (10)] |
| 25. | Chen C, Han Q, Ren H, Wu S, Li Y, Guo J, Li X, Liu X, Li C, Tian Y. Multiparametric MRI-based model for prediction of local progression of hepatocellular carcinoma after thermal ablation. Cancer Med. 2023;12:17529-17540. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 11] [Cited by in RCA: 8] [Article Influence: 2.7] [Reference Citation Analysis (0)] |
| 26. | Yang S, Sun B, Li J, Li N, Zhang A, Zhang X, Yang H, Zou X. Neutrophil extracellular traps promote angiogenesis in gastric cancer. Cell Commun Signal. 2023;21:176. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in RCA: 78] [Reference Citation Analysis (1)] |
| 27. | Yang LY, Luo Q, Lu L, Zhu WW, Sun HT, Wei R, Lin ZF, Wang XY, Wang CQ, Lu M, Jia HL, Chen JH, Zhang JB, Qin LX. Increased neutrophil extracellular traps promote metastasis potential of hepatocellular carcinoma via provoking tumorous inflammatory response. J Hematol Oncol. 2020;13:3. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 310] [Cited by in RCA: 317] [Article Influence: 52.8] [Reference Citation Analysis (19)] |
| 28. | Kaltenmeier CT, Yazdani H, van der Windt D, Molinari M, Geller D, Tsung A, Tohme S. Neutrophil extracellular traps as a novel biomarker to predict recurrence-free and overall survival in patients with primary hepatic malignancies. HPB (Oxford). 2021;23:309-320. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 15] [Cited by in RCA: 40] [Article Influence: 8.0] [Reference Citation Analysis (0)] |
| 29. | Wang H, Liu R, Mo H, Li R, Lian J, Liu Q, Han S. A novel nomogram predicting the early recurrence of hepatocellular carcinoma patients after R0 resection. Front Oncol. 2023;13:1133807. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 3] [Cited by in RCA: 15] [Article Influence: 5.0] [Reference Citation Analysis (0)] |
| 30. | Gu Y, Guo Y, Gao N, Fang Y, Xu C, Hu G, Guo M, Ma Y, Zhang Y, Zhou J, Luo Y, Zhang H, Wen Q, Qiao H. The proteomic characterization of the peritumor microenvironment in human hepatocellular carcinoma. Oncogene. 2022;41:2480-2491. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 1] [Cited by in RCA: 29] [Article Influence: 7.3] [Reference Citation Analysis (10)] |
| 31. | Yuan C, Wang Z, Gu D, Tian J, Zhao P, Wei J, Yang X, Hao X, Dong D, He N, Sun Y, Gao W, Feng J. Prediction early recurrence of hepatocellular carcinoma eligible for curative ablation using a Radiomics nomogram. Cancer Imaging. 2019;19:21. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 46] [Cited by in RCA: 75] [Article Influence: 10.7] [Reference Citation Analysis (11)] |
| 32. | Wang Y, Zhang Y, Xiao J, Geng X, Han L, Luo J. Multicenter Integration of MR Radiomics, Deep Learning, and Clinical Indicators for Predicting Hepatocellular Carcinoma Recurrence After Thermal Ablation. J Hepatocell Carcinoma. 2024;11:1861-1874. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 10] [Cited by in RCA: 10] [Article Influence: 5.0] [Reference Citation Analysis (9)] |
| 33. | Akinci D'Antonoli T, Farchione A, Lenkowicz J, Chiappetta M, Cicchetti G, Martino A, Ottavianelli A, Manfredi R, Margaritora S, Bonomo L, Valentini V, Larici AR. CT Radiomics Signature of Tumor and Peritumoral Lung Parenchyma to Predict Nonsmall Cell Lung Cancer Postsurgical Recurrence Risk. Acad Radiol. 2020;27:497-507. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 86] [Cited by in RCA: 75] [Article Influence: 12.5] [Reference Citation Analysis (0)] |
| 34. | Chong HH, Yang L, Sheng RF, Yu YL, Wu DJ, Rao SX, Yang C, Zeng MS. Multi-scale and multi-parametric radiomics of gadoxetate disodium-enhanced MRI predicts microvascular invasion and outcome in patients with solitary hepatocellular carcinoma ≤ 5 cm. Eur Radiol. 2021;31:4824-4838. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 183] [Cited by in RCA: 179] [Article Influence: 35.8] [Reference Citation Analysis (5)] |
| 35. | Wang QB, Luo WL, Li YK, Li J, Yang ZS, Zhao K, Lakang Y, Liang YB, Chen XM, Zuo JX, Duan Y, Xu X, Shang LM, Ke Y. Tumor Compression of the Hepatic or Portal Vein Predicts the Presence of Microvascular Invasion and Satellite Nodules in Hepatocellular Carcinoma: A Retrospective Study. J Hepatocell Carcinoma. 2025;12:2055-2067. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 23] [Cited by in RCA: 57] [Article Influence: 57.0] [Reference Citation Analysis (0)] |
| 36. | Liao M, Zhong X, Zhang J, Liu Y, Zhu Z, Wu H, Zeng Y, Huang J. Radiofrequency ablation using a 10-mm target margin for small hepatocellular carcinoma in patients with liver cirrhosis: A prospective randomized trial. J Surg Oncol. 2017;115:971-979. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 35] [Cited by in RCA: 39] [Article Influence: 4.3] [Reference Citation Analysis (3)] |
| 37. | Yu Y, Wang XH, Hu WJ, Chen DH, Hu ZL, Li SQ. Patterns, Risk Factors, and Outcomes of Recurrence After Hepatectomy for Hepatocellular Carcinoma with and without Microvascular Invasion. J Hepatocell Carcinoma. 2024;11:801-812. [RCA] [PubMed] [DOI] [Full Text] [Cited by in Crossref: 14] [Cited by in RCA: 16] [Article Influence: 8.0] [Reference Citation Analysis (0)] |
| 38. | Zhao Y, Wang S, Wang Y, Li J, Liu J, Liu Y, Ji H, Su W, Zhang Q, Song Q, Yao Y, Liu A. Deep learning radiomics based on contrast enhanced MRI for preoperatively predicting early recurrence in hepatocellular carcinoma after curative resection. Front Oncol. 2024;14:1446386. [RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)] [Cited by in Crossref: 10] [Cited by in RCA: 9] [Article Influence: 4.5] [Reference Citation Analysis (0)] |