This study was conducted in full compliance with the ethical principles of the Declaration of Helsinki. The Institutional Review Board granted formal approval for the study protocol. Due to the retrospective design, the requirement for individual patient consent was waived by the reviewing ethics committee.

This retrospective study evaluated 981 treatment-naïve HCC patients who underwent TACE between September 2015 and September 2024. Inclusion criteria consisted of: (1) age between 18 and 90 years; (2) adequate liver function (Child-Pugh class A or B); (3) Eastern Cooperative Oncology Group performance status of 0; and (4) pretreatment dynamic contrast-enhanced CT imaging obtained within two weeks prior to intervention. Exclusion criteria included: (1) uncorrectable coagulation disorders; (2) advanced hepatic insufficiency (Child-Pugh class C) or irreversible decompensation; (3) lack of TACE refractoriness; (4) insufficient follow-up documentation; (5) concurrent or previous malignant neoplasms; and (6) previous combination antitumor therapies. The patient enrollment process is illustrated in Fig. 1.

Patient enrollment flowchart. HCC hepatocellular carcinoma, TACE transarterial chemoembolization

Preoperative portal venous phase CE-CT images and baseline clinical data were retrospectively retrieved from the institutional electronic medical record system, including sex, age, Hepatitis B virus (HBV) infection status, Child-Pugh grade, liver function parameters, BCLC stage, number of tumors, largest tumor diameter, alpha-fetoprotein (AFP) levels, neutrophil-to-lymphocyte ratio (NLR), lymphocyte-to-monocyte ratio (LMR), and platelet-to-lymphocyte ratio (PLR).

This study was conducted following the abdominal CE-CT scanning protocol. Patients were instructed to fast from food and water for 4 to 6 h prior to the examination. Before scanning, patients underwent breathing training. The scan was performed in the supine position with arms elevated, and patients were instructed to hold their breath at the end of deep inspiration. The scanning range extended from the diaphragmatic dome to the lower pole of the kidney. A Discovery CT 750 HD (GE Healthcare) was utilized, with scanning parameters set as follows: tube voltage of 120 kV, automatic tube current modulation (noise index 12, GE CARE Dose4D technology), and real-time optimization of radiation dose based on patient body size. The reconstruction slice thickness was 5.0 mm. Optiray (320 mgI/mL, 1.5 ml/kg) was injected via the cubital vein at a flow rate of 3.0 ml/s, followed by a bolus injection of 20 ml of saline. Three-phase scans were performed at 25–30 s (arterial phase), 50–60 s (portal venous phase), and 90–120 s (delayed phase) post-injection.

The procedure was initiated using the Seldinger technique to puncture the right femoral artery, followed by contrast agent injection into the celiac trunk and common hepatic artery for angiography. This allowed visualization and confirmation of HCC location and its feeding arteries. Subsequently, a microcatheter was superselectively advanced into the tumor-feeding arteries, through which an appropriate mixture of epirubicin and lipiodol emulsion was delivered. For large hypervascular tumors, gelatin sponge particles were additionally employed to enhance the embolization effect [18].

According to the revised TACE resistance criteria by JSH-LCSGJ in 2021 [19], treatment resistance is not considered a termination indicator. The strategy should be adjusted based on tumor burden, liver function, and ECOG score (e.g., continuing TACE or combined therapy). Follow-up is conducted every 1 to 3 months postoperatively, with the endpoint established as February 1, 2025, or upon patient death. Overall survival (OS) is defined as the duration from the initial TACE to either death or the last follow-up.

Prior to segmentation, all patients’ portal venous phase CT images underwent standardized preprocessing using the open-source PyRadiomics package (Version 3.0.1) in Python, including voxel resampling to 1 mm × 1 mm × 1 mm isotropic resolution for spatial normalization and grayscale normalization to ensure intensity uniformity.

Target lesions were manually segmented by a board-certified radiologist (Observer A, 7 years of experience in abdominal imaging) using 3D Slicer software (version 5.6.2) to generate the intratumoral volume of interest (VOI). For patients with multifocal HCC, the lesion with the largest short-axis diameter was selected as the target lesion. This approach was implemented based on the following rationale: (1) the largest lesion is clinically designated as the “index lesion” and serves as a key reference for tumor staging and treatment strategy formulation; (2) it significantly reduces the workload and minimizes inter-observer variability associated with multi-lesion segmentation; (3) it aligns with established methodologies commonly employed in hepatocellular carcinoma radiomics research [20,21,22]. Three concentric peritumoral regions (extending 3 mm, 6 mm, and 10 mm beyond the tumor boundary) were automatically generated using SimpleITK’s 3D box kernel convolution. These regions were subsequently manually corrected by Observer A to exclude extrahepatic components, yielding the final Peri3mm, Peri6mm, and Peri10mm VOIs.

To assess interobserver variability, a second radiologist (Observer B, 10 years’ hepatobiliary imaging experience) independently segmented all VOIs in a randomly selected subset of 30 cases. Segmentation consistency was quantified using intraclass correlation coefficients (ICCs), with ICC > 0.75 considered indicative of excellent agreement.

Following the Image Biomarker Standardization Initiative (IBSI) guidelines, we extracted 1,834 radiomics features from each VOI (Intratumoral, Peri3mm, Peri6mm, and Peri10mm) using PyRadiomics in Python. The extracted features comprised seven categories: (1) first-order statistics, (2) gray-level co-occurrence matrix (GLCM), (3) gray-level dependence matrix (GLDM), (4) gray-level run length matrix (GLRLM), (5) gray-level size zone matrix (GLSZM), (6) neighboring gray tone difference matrix (NGTDM), and (7) shape features. All features underwent Z-score normalization for standardization. Feature selection was performed through a three-step process: (1) initial screening using Pearson correlation coefficient to remove highly redundant features (r > 0.9), (2) dimensionality reduction via least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation, (3) final selection using maximum relevance-minimum redundancy (mRMR) algorithm to identify optimal features for model construction.

The optimal retained radiomics features were utilized to construct models for Intratumoral, Peri3mm, Peri6mm, and Peri10mm regions using support vector machine (SVM), logistic regression (LR), and ExtraTrees (ET) algorithms, respectively. This process subsequently led to the development of a combined intratumoral-peritumoral model, achieved by integrating features from the optimal peritumoral region with those from the intratumoral region. Clinical predictors were identified through both univariate and multivariate logistic regression analyses. The final combined model incorporated both radiomics signatures and significant clinical predictors, and was presented as a nomogram for clinical application. The performance of the model was evaluated using receiver operating characteristic (ROC) curve analysis, which included the calculation of the area under the curve (AUC), accuracy (ACC), sensitivity (SENS), specificity (SPE), and F1-score. Decision curve analysis (DCA) was performed to evaluate the model’s goodness-of-fit and clinical net benefit. The maximum selected rank statistics method was used to identify the optimal cutoff value, and risk stratification was applied to assess patient prognosis. The workflow of the study is shown in Fig. 2.

All statistical analyses were performed using SPSS version 27.0 and Python programming language. Continuous variables underwent normality testing with the Kolmogorov-Smirnov method. Normally distributed data were expressed as mean ± standard deviation and compared using independent samples t-tests. Nonparametric variables were reported as median (interquartile range) and analyzed with Mann-Whitney U tests. Categorical variables were presented as counts (percentages) and examined using χ² tests. Survival outcomes were evaluated through Kaplan-Meier estimation, with between-group differences assessed by log-rank testing. A two-tailed p-value < 0.05 established statistical significance for all analyses.