Optimal 10-mm window: Integrating peri-ablation radiomics with preoperative features to predict early hepatocellular carcinoma recurrence after thermal ablation

Abstract

BACKGROUND

Hepatocellular carcinoma (HCC) remains a global health challenge. Thermal ablation (TA) is a guideline-recommended curative treatment for early-stage HCC. Early recurrence after TA, often arising from occult metastases missed during initial staging, is associated with a poorer prognosis. In this context, noninvasive imaging plays a critical role in assessing treatment response and stratifying recurrence risk in HCC.

AIM

To develop and validate a radiomics model integrating pre- and post-ablation features for predicting early HCC recurrence after TA.

METHODS

This retrospective study enrolled 301 patients who underwent initial TA for HCC and divided them into training (n = 168), internal validation (n = 72), and external validation (n = 61) sets. Radiomic features were extracted from pre-operative ultrasound and contrast-enhanced ultrasound images of the tumor, as well as from postoperative contrast-enhanced ultrasound images of four peri-ablation margins (5-20 mm from the ablation zone). Seven classifiers integrating clinical, pre-operative radiomic, and margin features were developed. Model performance was assessed using area under the curve (AUC), calibration, and decision curve analysis, followed by SHapley Additive exPlanations analysis.

RESULTS

A total of 301 patients, including 109 with ER, were included. The baseline model, based solely on pre-operative clinical and radiomic features, yielded an AUC of 0.849. Integration of features from concentric peri-ablation zones showed that model improvement depended on the spatial extent of the analyzed region. The optimal 10-mm peri-ablation support vector machine (SVM) model achieved the highest test-set AUC of 0.876 (95% confidence interval: 0.784-0.969), significantly outperforming the preoperative-only baseline model (AUC = 0.849). Models incorporating the 5-mm (AUC = 0.863) and 15-mm (AUC = 0.863) margins showed similar, though slightly lower, performance, whereas extending the analysis to a 20-mm margin resulted in a marked decline in performance (AUC = 0.795). SHapley Additive exPlanations analysis identified five top predictive radiomic features derived from pre-operative multiphase and postoperative portal-phase images.

CONCLUSION

Integrating pre-operative ultrasound with 10-mm peri-ablation radiomics effectively predicts early HCC recurrence after TA and may serve as a potential biomarker for risk stratification.

Key Words: Artificial intelligence; Hepatocellular carcinoma; Thermal ablation; Early recurrence; Contrast-enhanced ultrasound; Machine learning; Radiomics

Core Tip: Pre-operative ultrasound and contrast-enhanced ultrasound radiomic features can effectively predict early recurrence of hepatocellular carcinoma after thermal ablation. Imaging features of the 10-mm peri-ablation zone, when combined with pre-operative features, provide significant incremental predictive value and may help guide personalized adjuvant therapy and surveillance strategies.

INTRODUCTION

Hepatocellular carcinoma (HCC) accounts for 75%-85% of primary liver cancer cases. Owing to its increasing incidence and mortality over the past two decades, HCC continues to pose a serious global health challenge[1]. Advances in diagnostic imaging have enabled earlier detection of an increasing number of HCC cases[2]. Along with progress in interventional and minimally invasive techniques, thermal ablation (TA) has been established as a definitive, guideline-recommended curative treatment for early-stage HCC[3-6]. Radiofrequency ablation and microwave ablation, the two most common percutaneous ablation modalities, offer survival outcomes comparable to those of surgical resection[7-11]. However, the relatively high rate of early recurrence (ER) after ablation remains a major challenge[12,13]. ER in HCC, often resulting from occult metastases undetected during initial staging, portends a significantly poorer prognosis than in patients without ER[14]. Accurate identification of high-risk HCC patients who are likely to develop ER is critical. Imaging-based risk stratification can inform the use of neoadjuvant or adjuvant therapies, ultimately improving long-term survival[15].

Ultrasound and contrast-enhanced ultrasound (CEUS) are the primary and most important imaging modalities used throughout the perioperative management of HCC, from pre-operative decision-making and intraoperative guidance to postoperative surveillance, where their real-time capabilities offer unique advantages[16]. This established role makes them promising tools for predicting ER. A major challenge, however, is the subjective nature of image interpretation, which is inherently dependent on operator experience. Radiomics, an emerging imaging analysis technique, provides a quantitative and objective alternative to conventional visual assessment by extracting high-throughput features from digital images to characterize tumor heterogeneity, thereby offering a potential approach for predicting ER after ablation[17,18]. Previous studies have primarily explored this approach using computed tomography (CT) or magnetic resonance imaging (MRI)[19,20]. Although CT- and MRI-based radiomics models have been developed to predict HCC recurrence[21], ultrasound-based models, particularly those incorporating CEUS, remain markedly underdeveloped. Moreover, existing models have predominantly focused on the tumor itself while overlooking the surrounding postoperative ablation zone, the optimal spatial extent of which for ER prediction on ultrasound imaging has yet to be determined. This study aimed to develop a post-ablation predictive model integrating pre- and post-ablation ultrasound/CEUS images, clinical parameters, and serologic laboratory markers to predict early HCC recurrence after TA and support risk-stratified adjuvant therapy and surveillance.

MATERIALS AND METHODS

This dual-center retrospective study was conducted in accordance with the ethical standards of the Declaration of Helsinki. The study protocol was approved by the Institutional Ethics Committees of both participating institutions, Lanzhou University Second Hospital and Gansu Provincial Cancer Hospital. Informed consent was obtained from all participants or their legal guardians.

Study population and data source

Patients were retrospectively recruited from two independent medical centers: Lanzhou University Second Hospital (Center A) and Gansu Provincial Cancer Hospital (Center B). The dataset was chronologically divided into a training set (patients from Center A, January 2015 to December 2019), an internal validation set (patients from Center A, January 2020 to December 2023), and an external validation set (patients from Center B, January 2018 to June 2023) for model development and validation. The minimum follow-up duration was 2 years. The inclusion criteria were as follows: (1) HCC diagnosed by pathology or according to the noninvasive diagnostic criteria recommended by the European Association for the Study of the Liver; (2) A single tumor ≤ 5 cm or 2-3 tumors, each ≤ 3 cm in maximum diameter; (3) No evidence of vascular invasion or extrahepatic spread; (4) Complete clinical records and clear ultrasound/CEUS images obtained within 2 weeks before TA; and (5) Well-compensated liver disease.

The exclusion criteria were as follows: (1) Previous interventions for HCC; (2) Concurrent extrahepatic malignancies; (3) Unavailable laboratory or pathological data; (4) Loss to follow-up; (5) Suboptimal ultrasonographic visualization; (6) Severe dysfunction of other organ systems or contraindications to TA; (7) Incomplete ablation; and (8) Follow-up of < 6 months. Ultimately, 301 patients were enrolled in this study, of whom 109 developed ER and 192 did not. A flowchart of the study process is shown in Figure 1.

Figure 1 Study inclusion and exclusion flowchart. HCC: Hepatocellular carcinoma; TA: Thermal ablation; US: Ultrasound; CEUS: Contrast-enhanced ultrasound.

Treatment procedures

Following pretreatment CEUS localization, TA (either radiofrequency ablation or microwave ablation) was performed by experienced operators with more than 10 years of experience under real-time ultrasound guidance (65-80 W, 5-10 minutes). Single applicator placement was sufficient for tumors < 2 cm, whereas overlapping ablations were used for larger lesions to ensure a 5-mm margin. Protective measures, including artificial ascites or pleural effusion, were used as needed.

Acquisition of clinical data and ultrasound/CEUS images

Gray-scale ultrasound and CEUS were performed before ablation and again 1-2 months after ablation to evaluate the target lesion and ablation zone. Examinations were conducted using Philips EPIQ 7 or Siemens Sequoia systems equipped with C5-1 transducers (1-5 MHz). Following standard ultrasound, a 2.4 mL bolus of SonoVue contrast agent diluted in 5 mL of saline was administered intravenously, followed by a 5 mL saline flush. Both pre-operative and 1-2-month postoperative ultrasound and CEUS images were acquired and archived to minimize the confounding effects of tissue edema and acute inflammation that typically occur during the early post-ablation period. Clinical variables, including age, sex, etiology, tumor size, number of tumors, Barcelona Clinic Liver Cancer (BCLC) stage, and Child-Pugh grade, were collected. Laboratory parameters were retrieved from the electronic medical record system.

Follow-up and ER evaluation criteria

Patients were monitored at 1 month after TA, then every 3 months for 2 years, and every 6 months thereafter. Surveillance included serum alpha-fetoprotein (AFP), ultrasound, and contrast-enhanced CT or MRI. ER was defined as the detection of intrahepatic or extrahepatic recurrent lesions on follow-up imaging (contrast-enhanced CT, MRI, or CEUS) within 2 years after ablation[22-24]. For patients with contraindications to MRI, a combination of CT and CEUS was used.

Establishment process and methods of the multi-region radiomics model

The radiomics process flowchart is presented in Figure 2. Pre- and postoperative gray-scale and CEUS images were retrospectively collected, encompassing the complete target lesion, the background liver parenchyma, and the post-ablation zone. A total of 2709 ultrasonographic images, together with relevant clinical and laboratory data, were included in the analysis. The imaging dataset comprised pre-operative gray-scale images and both pre- and postoperative contrast-enhanced images acquired across multiple phases (arterial, early portal venous, late portal venous, and delayed). All images were systematically extracted from dynamic videos by radiologists with more than 20 years of clinical experience and were subsequently processed and securely managed on the Darwin scientific research platform (https://arxiv.org/abs/2009.00908) to ensure centralized, consistent handling. Clinical and laboratory variables were also imported into this platform.

Figure 2 Flowchart of the process from region of interest delineation on original ultrasound images to final prediction model construction. ROI: Region of interest; US: Ultrasound; CEUS: Contrast-enhanced ultrasound; SVM: Support vector machine; GBDT: Gradient boosting decision tree; XGBoost: Extreme gradient boosting; DT: Decision tree; RF: Random forest; LR: Logistic regression; KNN: K-nearest neighbors.

To mitigate inter-device heterogeneity and harmonize image contrast and brightness, all ultrasound images were normalized to a common gray-scale before region-of-interest (ROI) delineation, thereby ensuring consistent gray-scale representation across the dataset. ROI segmentation was performed as follows. Radiologist A (> 15 years of experience) and Radiologist B (> 10 years of experience) delineated the pre-operative tumor and postoperative ablation zone (Figure 3). Radiologist A repeated the segmentation twice at a 2-week interval [intra-observer intraclass correlation coefficient (ICC) = 0.92; 95% confidence interval (CI): 0.88-0.95], and Radiologist B performed it once independently (inter-observer ICC between A and B = 0.86; 95%CI: 0.80-0.91). Peri-ablation ROIs were then automatically generated by outward expansion from the ablation boundary using Darwin Scientific Research Platform v2.0. A series of ROIs representing the peripheral reactive zone surrounding the post-ablation necrotic area (peri-ablation zone) was automatically generated by outward expansion from the border of the necrotic zone, with expansion distances of 5, 10, 15, and 20 mm (Figure 4). When an automatically generated ROI extended beyond the skin contour, manual correction was performed by Radiologist C (> 10 years of experience) and Radiologist D (> 15 years of experience). Radiologist C repeated the correction twice at a 2-week interval (intra-observer ICC = 0.92; 95%CI: 0.89-0.95), and Radiologist D performed it once independently (inter-observer ICC between C and D = 0.88; 95%CI: 0.81-0.93). Throughout the process, the radiologists remained blinded to all baseline information and pathological findings, thereby ensuring objectivity and minimizing potential bias.

Figure 3 Pre-operative target lesion regions of interest were delineated and analyzed. A: Gray-scale ultrasound; B: The arterial phase of contrast-enhanced ultrasound; C: The early portal venous phase; D: The late portal venous phase; E: The delayed phase.

Figure 4 Delineation of the ablation zone margin and feature extraction from the peri-ablation region. Automated delineation of peri-ablation rims for feature extraction as yellow annuli. A: 5 mm peri-ablation rim; B: 10 mm peri-ablation rim; C: 15 mm peri-ablation rim; D: 20 mm peri-ablation rim.

After segmentation, radiomic features, including shape, first-order, and texture features, were extracted to characterize lesion geometry, voxel intensity distribution, and spatial heterogeneity, respectively. Both first-order and texture features were further transformed using six filters: Exponential, square, square root, logarithm, Laplacian of Gaussian (log-sigma-3-0 mm-3D), and wavelet. A total of 3272 pre-operative features and 1159 postoperative features were extracted per patient. For the integrated pre-operative + peri-necrotic 10-mm dataset, 4653 candidate variables entered the dimensionality-reduction pipeline. To minimize overfitting, preprocessing and feature selection were performed sequentially using min-max normalization, variance-based filtering, univariate F-test screening (P < 0.05), and recursive feature elimination. In the final optimal 10-mm model, 15 features were retained.

Model development

Seven machine learning classifiers were used to build predictive models for the following feature sets: (1) Pre-operative clinical features combined with radiomic features; and (2) Pre-operative clinical features combined with radiomic features further integrated with features from different peri-ablation zones (peri-necrotic 5 mm, peri-necrotic 10 mm, peri-necrotic 15 mm, and peri-necrotic 20 mm).

For each of the five feature sets, we trained and evaluated seven widely used supervised machine learning classifiers: K-nearest neighbors (KNN), SVM, logistic regression (LR), decision tree (DT), random forest (RF), gradient boosting DT (GBDT), and extreme gradient boosting (XGBoost). To improve generalizability, hyperparameters were calibrated for each algorithm. SVM used nonlinear kernels, LR used elastic net regularization, and all tree-based models were restricted to a maximum depth of 5.

Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score. All seven classifiers were assessed using 5-fold cross-validation within the training cohort. Importantly, all preprocessing and feature-selection steps were embedded within the cross-validation workflow. For each fold, min-max scaling, variance filtering, F-test screening, and recursive feature elimination were fitted exclusively on the analysis subset and then applied to the corresponding holdout fold. After model selection, the entire preprocessing-feature-selection-classifier pipeline was refitted on the full training cohort and then applied unchanged to the internal and external validation cohorts. Accordingly, no information from either validation cohort was used during feature selection or model tuning, thereby reducing the risk of information leakage and AUC inflation.

Statistical analysis

All analyses were conducted using SPSS 26.0 (IBM Corp, United States). Categorical variables are expressed as n (%). Continuous variables are presented as mean ± SD for normally distributed data and as median (interquartile range) for non-normally distributed data. Intergroup comparisons were performed using the independent-samples t-test for normally distributed data with homogeneous variance, the Mann-Whitney U test for non-normally distributed or heteroscedastic data, and the χ² test or Fisher’s exact test for categorical variables. Diagnostic performance across models was compared using DeLong’s test for differences in AUC. Calibration curves were plotted to evaluate agreement between predicted and observed outcomes, and clinical utility was assessed using decision curve analysis (DCA) to quantify net benefit.

RESULTS

Demographic and clinical characteristics of patients

A total of 301 patients were included in this study, comprising 109 patients with ER and 192 without ER. Of these, 168 patients (67 with ER and 101 without ER) were assigned to the training set, 72 (20 with ER and 52 without ER) to the internal validation set, and 61 (22 with ER and 39 without ER) to the external validation set. The clinical characteristics of the ER and non-ER groups are summarized in Table 1.

Table 1 Characteristics of the included patients, n (%)/mean ± SD.

	Training set			P value	Internal validation set (n = 72)	P value	External validation set (n = 61)	P value
	Total (n = 168)	ER (n = 67)	Non-ER (n = 101)
Patient characteristic
Age (years)	56.54 ± 9.56	56.15 ± 9.73	56.79 ± 9.48	0.671	56.28 ± 8.73	0.84	54.95 ± 9.34	0.266
Sex				0.874		0.215		0.173
Male	123 (73.2)	48 (71.6)	75 (74.3)		47 (65.3)		50 (65.3)
Female	45 (26.8)	19 (28.4)	26 (25.7)		25 (34.7)		11 (34.7)
Etiology				0.147		0.915		0.934
HBV	155 (92.3)	64 (95.5)	91 (90.1)		67 (93.1)		57 (93.4)
HCV	9 (5.4)	3 (4.5)	6 (5.9)		3 (4.2)		3 (4.9)
Others	4 (2.4)	0 (0.0)	4 (4.0)		2 (2.8)		1 (1.6)
Tumor characteristics
Tumor size (cm)	2.39 ± 0.91	2.72 ± 0.91	2.18 ± 0.91	< 0.001	2.33 ± 0.81	0.651	2.46 ± 0.97	0.612
No. of tumors				< 0.001		0.705		0.516
Solitary	123 (73.2)	39 (58.2)	84 (83.2)		51 (70.8)		42 (68.9)
Multiple	45 (26.8)	28 (41.8)	17 (16.8)		21 (29.2)		19 (31.1)
Barcelona Clinic Liver Cancer stage				< 0.001		0.702		0.862
0	46 (27.4)	5 (7.5)	41 (40.6)		18 (25.0)		16 (26.2)
A	122 (72.6)	62 (92.5)	60 (59.4)		54 (75.0)		45 (73.8)
Laboratory index
AST, IU/L	36.86 ± 5.25	37.00 ± 5.38	36.76 ± 5.19	0.779	36.25 ± 5.22	0.408	35.19 ± 5.41	0.036
ALT, IU/L	31.01 ± 21.78	34.42 ± 23.96	28.76 ± 20.01	0.099	33.12 ± 25.71	0.517	28.23 ± 15.93	0.362
TBIL, μmol/L	24.59 ± 13.65	25.63 ± 14.80	23.89 ± 12.87	0.423	21.34 ± 11.74	0.08	21.44 ± 12.96	0.12
ALB, g/L	30.78 ± 12.66	31.36 ± 10.08	30.40 ± 14.14	0.634	33.43 ± 20.90	0.229	32.74 ± 10.53	0.28
r-GGT (U/L)	35.78 ± 26.68	38.60 ± 31.92	33.91 ± 22.52	0.266	32.01 ± 18.89	0.278	40.42 ± 29.96	0.261
PLT, × 109/L	110.54 ± 55.67	106.62 ± 51.70	113.14 ± 58.14	0.459	112.81 ± 56.08	0.774	104.41 ± 54.45	0.459
Child-Pugh grade				0.249		0.975		0.204
A	121 (72.0)	47 (70.1)	74 (73.3)		52 (72.2)		49 (80.3)
B	47 (29.0)	20 (29.9)	27 (26.7)		20 (27.8)		12 (19.7)
AFP, ng/mL	215.98 ± 351.69	329.83 ± 440.58	140.44 ± 252.97	0.001	187.27 ± 331.25	0.556	294.55 ± 378.51	0.144

ER: Early recurrence; HBV: Hepatitis B virus; HCV: Hepatitis C virus; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; TBIL: Total bilirubin; ALB: Albumin; r-GGT: Gamma-glutamyl transferase; PLT: Platelet; AFP: Alpha-fetoprotein.

No significant differences were observed between the ER and non-ER groups with respect to age (P = 0.671), sex (P = 0.874), etiology (P = 0.147), Child-Pugh grade (P = 0.249), aspartate aminotransferase level (P = 0.779), alanine aminotransferase (ALT) level (P = 0.099), total bilirubin level (P = 0.423), albumin level (P = 0.634), gamma-glutamyl transferase level (P = 0.266), or platelet count (P = 0.459). In contrast, the number of tumors (P < 0.001), tumor size (P < 0.001), BCLC stage (P < 0.001), and AFP level (P = 0.001) were significantly higher in patients with ER than in those without ER. The external validation set had a significantly higher ALT level than the other two sets (P = 0.036). Aside from this difference, the three datasets were similar in all other baseline clinical characteristics (all P > 0.05).

Development of pre-operative models

Initially, we sought to establish a baseline predictive model using only information available before treatment. To this end, we constructed a comprehensive pre-operative feature set by integrating conventional clinical variables with radiomic features extracted from pre-ablation imaging. Seven distinct machine learning classifiers were trained and optimized using 5-fold cross-validation on this pre-operative dataset. Among the seven classifiers trained on the pre-operative feature set, SVM demonstrated the highest discriminative ability in the external validation set, with an AUC of 0.849 (95%CI: 0.747-0.951) (Table 2). This pre-operative SVM model was therefore established as the baseline for subsequent comparisons.

Table 2 Predictive performance of seven machine learning models of all classifiers across different feature sets, area under the curve (95% confidence interval).

		KNN	SVM	LR	DT	GBDT	RF	XGBoost
Clinical + rad-preop	Training set	0.966 (0.947-0.984)	0.928 (0.893-0.964)	0.941 (0.91-0.973)	0.757 (0.701-0.814)	0.973 (0.954-0.992)	0.909 (0.871-0.946)	0.987 (0.974-1)
	Internal validation set	0.76 (0.631-0.888)	0.848 (0.936-0.98)	0.867 (0.772-0.963)	0.756 (0.64-0.87)	0.845 (0.799-0.892)	0.76 (0.636-0.884)	0.864 (0.817-0.91)
	External validation set	0.752 (0.622-0.883)	0.849 (0.747-0.951)	0.84 (0.735-0.945)	0.716 (0.597-0.836)	0.825 (0.715-0.935)	0.752 (0.625-0.878)	0.83 (0.72-0.939)
Clinical + rad-preop + peri-necrotic (5 mm)	Training set	0.97 (0.954-0.987)	0.957 (0.933-0.982)	0.963 (0.94-0.986)	0.757 (0.701-0.814)	0.986 (0.974-0.999)	0.917 (0.88-0.953)	0.993 (0.984-1)
	Internal validation set	0.938 (0.911-0.964)	0.894 (0.852-0.935)	0.865 (0.819-0.911)	0.751 (0.618-0.883)	0.933 (0.898-0.967)	0.839 (0.788-0.89)	0.839 (0.928-0.985)
	External validation set	0.81 (0.692-0.928)	0.863 (0.761-0.965)	0.859 (0.752-0.966)	0.716 (0.597-0.836)	0.819 (0.706-0.932)	0.824 (0.709-0.94)	0.816 (0.702-0.93)
Clinical + rad-preop + peri-necrotic (10 mm)	Training set	0.952 (0.929-0.975)	0.942 (0.912-0.972)	0.95 (0.921-0.979)	0.757 (0.701-0.814)	0.973 (0.953-0.993)	0.913 (0.875-0.951)	0.987 (0.975-1)
	Internal validation set	0.833 (0.717-0.949)	0.898 (0.859-0.936)	0.898 (0.859-0.937)	0.736 (0.678-0.794)	0.907 (0.878-0.943)	0.803 (0.688-0.919)	0.904 (0.869-0.939)
	External validation set	0.818 (0.7-0.935)	0.876 (0.784-0.969)	0.873 (0.775-0.971)	0.716 (0.597-0.836)	0.871 (0.773-0.969)	0.795 (0.674-0.917)	0.855 (0.749-0.961)
Clinical + rad-preop + peri-necrotic (15 mm)	Training set	0.96 (0.94-0.98)	0.941 (0.908-0.974)	0.95 (0.919-0.981)	0.757 (0.701-0.814)	0.976 (0.958-0.994)	0.913 (0.876-0.95)	0.985 (0.97-1)
	Internal validation set	0.927 (0.893-0.96)	0.875 (0.784-0.967)	0.873 (0.777-0.9969)	0.725 (0.599-0.851)	0.913 (0.876-0.951)	0.833 (0.72-0.946)	0.839 (0.74-0.937)
	External validation set	0.851 (0.747-0.954)	0.863 (0.763-0.962)	0.856 (0.754-0.958)	0.716 (0.597-0.836)	0.84 (0.734-0.946)	0.815 (0.7-0.93)	0.833 (0.727-0.939)
Clinical + rad-preop + peri-necrotic (20 mm)	Training set	0.971 (0.955-0.987)	0.986 (0.974-0.997)	0.958 (0.934-0.982)	0.767 (0.711-0.823)	0.927 (0895-0.959)	0.918 (0.885-0.952)	0.995 (0.99-1)
	Internal validation set	0.771 (0.712-0.831)	0.792 (0.67-0.913)	0.79 (0.732-0.849)	0.742 (0.684-0.8)	0.801 (0.676-0.926)	0.81 (0.691-0.929)	0.834 (0.72-0.948)
	External validation set	0.717 (0.583-0.851)	0.795 (0.676-0.915)	0.773 (0.645-0.9)	0.664 (0.537-0.791)	0.763 (0.637-0.889)	0.753 (0.625-0.881)	0.782 (0.654-0.91)

KNN: K-nearest neighbors; SVM: Support vector machine; LR: Logistic regression; DT: Decision tree; GBDT: Gradient boosting decision tree; RF: Random forest; XGBoost: Extreme gradient boosting.

Construction of post-ablation feature sets

As an exploratory step to investigate the incremental prognostic value of peri-ablation zone characteristics, we developed four sequential models by augmenting the baseline pre-operative feature set with radiomic features derived from the peri-ablation zone. Specifically, we defined and extracted features from concentric margins 5 mm, 10 mm, 15 mm, and 20 mm surrounding the ablation necrotic zone. These features were combined with the pre-operative features to generate four new composite datasets: Preop + peri-necrotic 5 mm, preop + peri-necrotic 10 mm, preop + peri-necrotic 15 mm, and preop + peri-necrotic 20 mm. Each of these four composite datasets was then used to train the same panel of seven machine learning classifiers following an identical model development and validation protocol (Table 2). The incorporation of peri-ablation features was associated with improved predictive performance, except for the 20-mm margin. The magnitude of performance improvement depended on the selected margin distance.

The model incorporating the 10-mm peri-necrotic margin demonstrated the highest discriminatory ability. Specifically, the best-performing SVM classifier based on this margin achieved a peak AUC of 0.876 (95%CI: 0.784-0.969) in the independent validation set. This performance was significantly superior to that of the baseline pre-operative model (AUC = 0.849; 95%CI: 0.747-0.951; P = 0.02) and to the SVM models based on other peri-necrotic margins (5 mm: AUC = 0.863, P = 0.04; 15 mm: AUC = 0.863, P = 0.04; 20 mm: AUC = 0.795, P = 0.01; Table 3, Figure 5).

Figure 5 Comparison of receiver operating characteristic curves for different models. A: Training set; B: Internal validation set; C: External validation set. AUC: Area under the curve.

Table 3 Comparative performance of the optimal preoperative clinical-radiomic model vs optimal models incorporating peri-ablation zones features, area under the curve (95% confidence interval).

		AUC	Sensitivity	Specificity	PPV	NPV	Accuracy	F1 score
Clinical + rad-preop	Training set	0.928 (0.893-0.964)	0.934 (0.882-0.964)	0.809 (0.715-0.877)	0.764 (0.713-0.832)	0.949 (0.803-0.968)	0.859	0.841
	Internal validation set	0.848 (0.742-0.955)	0.842 (0.696-0.926)	0.739 (0.535-0.875)	0.554 (0.503-0.773)	0.924 (0.835-0.975)	0.768	0.668
	External validation set	0.849 (0.747-0.951)	0.763 (0.608-0.87)	0.826 (0.629-0.93)	0.707 (0.698-0.847)	0.864 (0.747-0.897)	0.804	0.734
Clinical + rad-preop + peri-necrotic (5 mm)	Training set	0.957 (0.933-0.982)	0.921 (0.866-0.954)	0.888 (0.805-0.938)	0.845 (0.791-0.903)	0.944 (0.803-0.965)	0.901	0.881
	Internal validation set	0.894 (0.852-0.9935)	0.821 (0.752-0.874)	0.831 (0.74-0.895)	0.651 (0.613-0.734)	0.924 (0.839-0.979)	0.828	0.726
	External validation set	0.863 (0.761-0.965)	0.816 (0.666-0.908)	0.783 (0.581-0.903)	0.674 (0.615-0.839)	0.886 (0.724-0.927)	0.803	0.795
Clinical + rad-preop + peri-necrotic (10 mm)	Training set	0.942 (0.912-0.972)	0.894 (0.835-0.934)	0.854 (0.766-0.913)	0.803 (0.782-0.901)	0.924 (0.836-0.974)	0.87	0.846
	Internal validation set	0.898 (0.859-0.936)	0.815 (0.745-0.868)	0.843 (0.753-0.904)	0.666 (0.612-0.788)	0.922 (0.874-0.965)	0.835	0.733
	External validation set	0.876 (0.784-0.969)	0.947 (0.827-0.985)	0.652 (0.449-0.812)	0.59 (0.543-0.705)	0.957 (0.857-0.997)	0.757	0.734
Clinical + rad-preop + peri-necrotic (15 mm)	Training set	0.941 (0.908-0.974)	0.907 (0.85-0.944)	0.854 (0.766-0.913)	0.805 (0.743-0.849)	0.933 (0.856-0.969)	0.875	0.853
	Internal validation set	0.875 (0.784-0.967)	0.789 (0.637-0.889)	0.826 (0.629-0.93)	0.636 (0.612-0.782)	0.911 (0.815-0.941)	0.816	0.704
	External validation set	0.863 (0.763-0.962)	0.868 (0.727-0.942)	0.783 (0.581-0.903)	0.688 (0.627-0.742)	0.915 (0.831-0.958)	0.813	0.768
Clinical + rad-preop + peri-necrotic (20 mm)	Training set	0.986 (0.974-0.997)	0.94 (0.891-0.968)	0.955 (0.889-0.982)	0.933 (0.867-0.989)	0.960 (0.826-0.987)	0.949	0.937
	Internal validation set	0.792 (0.67-0.913)	0.868 (0.727-0.942)	0.696 (0.491-0.844)	0.523 (0.501-0.713)	0.932 (0.849-0.987)	0.744	0.653
	External validation set	0.795 (0.676-0.915)	0.658 (0.499-0.788)	0.864 (0.667-0.953)	0.727 (0.701-0.863)	0.821 (0.723-0.895)	0.791	0.691

Data are 95% confidence intervals in parentheses.

AUC: Area under the curve; PPV: Positive predictive value; NPV: Negative predictive value.

Figure 6 shows the receiver operating characteristic curves for the seven machine learning models developed using the pre-operative + peri-necrotic 10-mm feature set. In the external validation set, the AUC values for the KNN, SVM, LR, DT, GBDT, RF, and XGBoost classifiers were 0.818 (95%CI: 0.700-0.935), 0.876 (95%CI: 0.784-0.969), 0.873 (95%CI: 0.775-0.971), 0.716 (95%CI: 0.597-0.836), 0.871 (95%CI: 0.773-0.969), 0.795 (95%CI: 0.674-0.917), and 0.855 (95%CI: 0.749-0.961), respectively.

Figure 6 Performance comparison on the external validation set of the seven machine learning models based on the “pre-operative + peri-necrotic 10-mm” feature set, using receiver operating characteristic curves. ROC: Receiver operating characteristic; XGBoost: Extreme gradient boosting; RF: Random forest; GBDT: Gradient boosting decision tree; SVM: Support vector machine; DT: Decision tree; LR: Logistic regression; KNN: K-nearest neighbors; AUC: Area under the curve.

Results of DeLong’s test indicated no significant differences in AUC among the SVM, GBDT, LR, and XGBoost models (all P > 0.05). RF showed performance comparable to that of GBDT and XGBoost (P > 0.05) but was significantly outperformed by SVM and LR (P < 0.05). In contrast, DT performed significantly worse than SVM, LR, and RF (all P < 0.05), and KNN was significantly outperformed by both SVM and LR (P < 0.05). Among all models, SVM achieved the highest AUC in the external validation set, reaching 0.876 (95%CI: 0.784-0.969).

Figure 7 presents the DCA. The DCA demonstrated the favorable clinical utility of the pre-operative + peri-necrotic 10-mm SVM model across a reasonable range of threshold probabilities. Figure 8 presents the calibration curves. Both the training and external validation sets showed excellent calibration, indicating strong agreement between predicted and observed probabilities.

Figure 7 Decision curve analysis of the support vector machine model developed using the “pre-operative + peri-necrotic 10-mm” feature set. A: Decision curves on the training set; B: Decision curves on the internal validation set; C: Decision curves on the external validation set.

Figure 8 Model calibration was evaluated in the training and external validation set. SVM: Support vector machine.

Visualization of the optimal model

We performed SHapley Additive exPlanations (SHAP) analysis to interpret the optimal pre-operative + peri-necrotic 10-mm SVM model. SHAP analysis provides both local and global interpretability by quantifying each feature’s contribution to individual predictions and capturing potential nonlinear interactions. Figure 9A shows the 15 retained features in the final SVM model. The five highest-contributing features were wavelet-HH_glcm_SumEntropy_Pre-US, original_shape2D_MinorAxisLength_Pre-AP, wavelet-HH_glszm_SizeZoneNonUniformityNormalized_Pre-EPVP, lbp-2D_gldm_DependenceEntropy_POST-PVP_Peri-tumoral, and wavelet-HL_ngtdm_Complexity_Pre-AP. Figure 9B ranks the 15 retained features by their mean absolute SHAP values, illustrating their relative contributions to the model output. Positive SHAP values (to the right of the vertical line) indicate a positive impact on the prediction. Red indicates higher feature values, whereas blue indicates lower feature values.

Figure 9 Feature selection and visualization of radiomics showed the “pre-operative + peri-necrotic 10-mm” model’s interpretation. A: The importance ranking of the 15 factors based on the support vector machine classifier; B: SHapley Additive exPlanations values showed the model interpretation, the red part represented the higher values. SHAP: SHapley Additive exPlanations.

DISCUSSION

Despite favorable safety profiles and survival outcomes comparable to those of surgery[25], predicting ER after TA for HCC remains challenging. Because ER is strongly associated with poor prognosis[26], early identification of high-risk patients is essential for guiding treatment decisions and personalizing surveillance strategies.

In our study, patients with ER exhibited significantly greater tumor burden, reflected by both larger tumor size and a higher number of tumors, more advanced BCLC stage, and higher serum AFP levels than those without ER. These findings are consistent with established prognostic frameworks for HCC. First, larger tumor size is a well-recognized factor that increases procedural complexity and reduces the likelihood of curative ablation, as supported by previous studies[27]. Second, the presence of multiple tumors, which was a hallmark of ER in our cohort, may reflect underlying multicentric oncogenesis, a pattern frequently associated with imaging-occult disease and a greater propensity for metastasis. Our study also showed that patients classified as BCLC stage A had a significantly higher rate of ER than those with BCLC stage 0 disease. This difference was statistically significant (P < 0.001), suggesting that finer stratification within early-stage HCC carries distinct prognostic implications for treatment outcomes, a finding consistent with previous reports[28]. Finally, our results corroborate previous reports[29] identifying elevated AFP as a robust serologic marker of increased risk of ER following locoregional therapy.

Artificial intelligence excels at identifying predictive patterns beyond human perception[30]. Radiomics transforms medical images into high-dimensional quantitative features. By integrating clinical and imaging data end-to-end, artificial intelligence-based radiomics enables automated feature extraction and outcome prediction, thereby addressing the limitations of conventional models[31,32]. Existing prediction models for HCC recurrence after ablation have primarily relied on pre-operative clinical and imaging data[33,34] and have not incorporated characteristics of the peri-ablation zone. To address this gap, we developed a predictive model for ER by integrating pre-operative radiomics with features from the postoperative peri-necrotic zone in a cohort of 301 patients with HCC who underwent initial TA. Notably, a predictive model based solely on preprocedural information, integrating standard clinical parameters with radiomic features from pre-ablation imaging, achieved robust performance (AUC = 0.849) in predicting ER. To investigate the incremental prognostic value of peri-ablation zone characteristics, we developed four sequential models (preop + peri-necrotic 5 mm, preop + peri-necrotic 10 mm, preop + peri-necrotic 15 mm, and preop + peri-necrotic 20 mm) by augmenting the baseline pre-operative feature set with radiomic features from the peri-ablation zone. Notably, integration of peri-ablation radiomic features was consistently associated with improved predictive performance. However, the degree of improvement was not monotonic with respect to the spatial extent of the analyzed postoperative region. We observed diminishing returns beyond a 10 mm boundary, with performance declining at a 20 mm margin. This pattern suggests the presence of a spatially confined, biologically relevant high-risk rim.

Conversely, the multiparametric model that incorporated pre-operative clinical variables, tumor radiomics, and features from the 10 mm peri-ablation zone demonstrated optimal performance (AUC = 0.876), significantly outperforming the pre-operative model. Although the model using the 20-mm margin showed near-perfect performance on the training set (AUC = 0.986), it exhibited a marked decline on the external validation set (AUC = 0.795), indicating substantial overfitting. In contrast, the 10-mm model achieved an optimal balance, with excellent and stable discriminative ability in both the training set (AUC = 0.942) and, more importantly, the independent validation set (AUC = 0.876). This consistency supports its superior generalizability and identifies it as the optimal model for clinical prediction.

Recurrence after ablation may be attributed to insufficient TA and the associated sterile inflammatory response surrounding the coagulative necrosis area[35]. The peri-ablation zone, adjacent to the necrotic core, sustains sublethal thermal injury, where M2 macrophages promote tissue repair through matrix remodeling, angiogenesis, and immunosuppression[36]. Increased M2 macrophage activity promotes residual tumor growth and contributes to unfavorable clinical outcomes in patients with HCC[37]. Consequently, imaging features derived from this region, such as textural heterogeneity and perfusion variation, may reflect the inflammatory-repair microenvironment and serve as biomarkers of treatment response. For example, an irregular peri-ablation rim may suggest incomplete ablation or microscopic residual disease. Thus, peri-ablation imaging phenotypes may indicate both the local biological response and the risk of recurrence.

Feature-importance analysis identified five dominant radiomic signatures, further supporting the biological relevance of multimodal integration. The leading feature, wavelet-HH_glcm_SumEntropy_Pre-US, reflects gray-scale textural disorder on pre-operative conventional ultrasound. Higher entropy generally indicates greater intratumoral heterogeneity, which may be associated with aggressive histologic architecture, microvascular invasion, or microscopic necrosis.

The second feature, original_shape2D_MinorAxisLength_Pre-AP, captures lesion geometry on pre-operative arterial-phase CEUS and may reflect morphologic compactness or irregular expansion patterns associated with infiltrative growth. The third feature, wavelet-HH_glszm_SizeZoneNonUniformityNormalized_Pre-EPVP, quantifies heterogeneity in the size distribution of homogeneous intensity zones during the pre-operative early portal venous phase, thereby characterizing spatially uneven washout behavior. The fourth feature, lbp-2D_gldm_DependenceEntropy_POST-PVP_Peri-tumoral, describes the randomness of local gray-level dependence within the peri-ablation rim on postoperative portal venous-phase CEUS and may reflect a heterogeneous inflammatory-repair interface, microcirculatory disturbance, or residual viable microscopic foci. Finally, wavelet-HL_ngtdm_Complexity_Pre-AP captures local intensity variation during the pre-operative arterial phase and may be associated with disordered neoangiogenesis and perfusion heterogeneity. Taken together, these findings suggest that the risk of ER is jointly encoded by pretreatment tumor aggressiveness and the biological response of the peri-ablation microenvironment rather than by a single imaging domain.

Among the seven classifiers evaluated, SVM consistently demonstrated the most robust overall performance across feature sets. This likely reflects the suitability of SVM for relatively small-sample, high-dimensional radiomics data, in which complex but smooth nonlinear boundaries can be modeled through kernel transformation while maintaining stronger regularization than many tree-based learners. By contrast, several tree-based models, particularly XGBoost and GBDT, achieved near-perfect discrimination in the training cohort but showed larger declines in the validation cohorts. In this dataset, that pattern is most consistent with overfitting caused by the combination of limited event numbers, high-dimensional candidate features, and the presence of noisy or weakly informative variables, especially when the peri-ablation window was expanded to 20 mm.

Extending the margin to 20 mm may not only dilute the biologically relevant peri-ablation signal by incorporating more normal liver parenchyma but also introduce additional high-dimensional noise that tree-based models may overfit during training. Accordingly, the 20-mm model appears to reflect both signal dilution and noise-driven overfitting. In contrast, the 10-mm SVM model achieved a better balance between discriminative ability and generalizability, supporting its selection as the final model.

Although the external validation cohort (Center B) had significantly higher ALT levels than the training and internal validation cohorts (P = 0.036, Table 1), suggesting inter-center heterogeneity, this difference is unlikely to have substantially affected model performance. ALT is a marker of hepatocellular injury rather than a direct determinant of tumor biology or CEUS-derived radiomic features. Furthermore, ALT has not been established as an independent predictor of ER after TA in HCC. Importantly, our model achieved robust performance in the external validation set (AUC = 0.876), comparable to that in the internal validation set (AUC = 0.898), indicating that inter-center heterogeneity in ALT did not meaningfully impair generalizability.

Recent studies have validated the utility of ultrasound-based radiomics for predicting HCC prognosis. For example, Jiang et al[38] developed an ultrasound-based radiomics model for ER after resection (AUC = 0.884), confirming the predictive value of radiomic features. Similarly, Ma et al[33] integrated gray-scale ultrasound features with machine learning and demonstrated that imaging heterogeneity is associated with aggressive tumor behavior (AUC = 0.84). However, our study differs from and extends these findings in several respects. First, whereas most existing models focus on surgical resection, we examined TA, a setting in which pathologic margin assessment is unavailable, making the peri-ablation zone uniquely informative. Second, previous radiomics analyses have primarily extracted features from the primary tumor, whereas we incorporated features from post-ablation peri-necrotic regions. The finding that GLDM features from this region predict recurrence supports the hypothesis that microscopic residual disease within the transition zone drives ER. Third, our multiphase approach, incorporating pre-operative gray-scale ultrasound, CEUS, and post-ablation imaging, provides a more dynamic assessment of tumor biology and treatment response.

In the independent validation set, our 10-mm model achieved a sensitivity of 81.6% (95%CI: 66.6%-90.8%) and a specificity of 78.3% (95%CI: 58.1%-90.3%). This performance profile suggests a tendency toward over-prediction of ER. In clinical practice, such false-positive predictions may lead to unnecessary adjuvant therapy. However, the model is better suited as a risk-stratification screening tool and therefore prioritizes sensitivity to avoid missing true ERs. Future studies should aim to improve specificity further by integrating complementary biomarkers or optimizing radiomic feature selection.

From a clinical implementation perspective, this model could function as an ultrasound-based decision-support tool at two practical time points: Before ablation, to identify patients at higher risk of ER and refine treatment planning; and after ablation, to tailor surveillance intensity. Because gray-scale ultrasound and CEUS are already embedded in routine HCC management, future clinical translation may be feasible through workflow-integrated software that imports DICOM images or cine loops, automatically performs region segmentation and feature extraction, and outputs an individualized recurrence-risk score with minimal disruption to clinical workflow.

Our study has several limitations. First, because of the retrospective design and the exclusion of a very small number of cases with ER due to technical issues during the ablation procedure, the model’s generalizability requires further validation in prospective, multicenter cohorts. Second, the ultrasound images used were acquired from diverse scanners, which, while reflecting real-world practice, may affect the model’s generalizability to new imaging sources. Third, the lack of prospective cohort validation prevents confirmation of the actual benefits of adjuvant chemotherapy in high-risk patients. Fourth, the peri-ablation zone was delineated on CEUS images from a single time point (1-2 months post-ablation), without accounting for the dynamic inflammatory and vascular changes in this region. Future studies using delta-radiomics to track the temporal evolution of radiomic features are warranted to improve prediction.

CONCLUSION

We developed an optimal model integrating pre-operative clinical-radiomic features with a 10-mm peri-ablation radiomic signature. This model outperformed both the preoperative-only model and models incorporating other peri-ablation margins, thereby providing early risk stratification to support personalized surveillance and adjuvant therapy, potentially improving patient outcomes.