IntroductionHepatocellular carcinoma (HCC) is one of the most common malignancies and the leading cause of cancer-related mortality worldwide1. A significant proportion of patients are diagnosed at an advanced stage, losing the opportunity for curative surgical resection and resulting in a poor prognosis2.Systemic treatments with molecular and immune therapies have dramatically transformed the management landscape of advanced HCC. IMbrave-150, ORIENT-32 and CARES-310 trials have demonstrated that the combination of molecular targeted therapy with immunotherapy can improve the prognosis of patients with advanced HCC effectively and safely3,4,5. Now immune checkpoint inhibitors (ICIs) plus molecular targeted therapies (MTTs) have been recommended as the first-line treatment for advanced HCC6. Transcatheter arterial chemoembolization (TACE) is recommended as a standard therapy for intermediate-stage HCC and widely used in patients with unresectable HCC in real-world practice. Studies hold views that TACE can modulate tumor immune microenvironment, and its combination with ICIs and MTTs may exert a synergistic effect7. Moreover, data from trials like EMERALD-1, CHANCE have shown that TACE-ICIs-MTTs regimen demonstrated promising results, with significantly longer overall survival (OS) and progression-free survival (PFS)8,9. However, given the biological complexity of HCC and its heterogeneity across individuals, many patients do not derive a survival benefit from treatment while simultaneously facing the risk of severe adverse effects and financial burden. There is a lack of effective predictive tools to identify patients who are more likely to benefit from TACE-ICIs-MTTs.Radiomics can extract high-throughput quantitative features from clinical imaging and provide a wealth of pathological and biological characteristics information about disease that are difficult to distinguish with the naked eye, holds significant potential in advancing precision medicine in a non-invasive and reproducible way. Currently, radiomics has been widely used in the investigation of issues related to the diagnosis, pathological characteristic evaluation, and prediction of therapeutic efficacy in HCC10,11,12. A number of radiomics models with favorable performance have been established to predict the treatment response of TACE-ICIs-MTTs13,14,15. These findings underscore its potential as an effective predictive tool. However, there are relatively few studies focusing on predicting the prognosis of TACE-ICIs-MTTs.The aim of this study was to develop and validate a contrast-enhanced CT (CECT)-based clinical-radiomics model for predicting the overall survival (OS) in unresectable HCC patients who received TACE-ICIs-MTTs.MethodsStudy participantsThis study was approved by the institutional review board, which waived the requirement for informed consent due to its retrospective nature. This multicenter retrospective study enrolled 180 HCC patients meeting the inclusion and exclusion criteria from January 2020 to December 2024 in four medical centers: W Hospital (center 1), X Hospital (center 2), Y Hospital (center 3), Z Hospital (center 4). Inclusion criteria: (1) patients age > 18 years; (2) patients with clinically or pathologically confirmed diagnosis of HCC according to the Chinese Guidelines for Diagnosis and Treatment of Primary Liver Cancer16; (3) patients who received TACE combined with a PD-(L)1 inhibitor and either an anti-VEGF antibody or a tyrosine kinase inhibitor (TKI) as first-line treatment; (4) patients who had undergone a contrast-enhanced CT scan within 2 weeks prior to treatment; (5) liver function Child-Pugh Class A or B and (6) Eastern Cooperative Oncology Group (ECOG) performance status of 0–2. Exclusion criteria: (a) prior surgery, ablation, targeted therapy, or immunotherapy before the baseline CT scans; (b) a history of other malignant tumor; (c) incomplete clinical records; (d) absence of CT images or poor image quality; and (e) lost to follow-up.The study comprised of 3 cohorts, with 132 patients from center 1 were randomly assigned to the training cohort (n = 93) and the internal validation cohort (n = 39), following a 7:3 ratio. The external validation cohort included 48 patients from center 2–4 (Fig. 1).Fig. 1Full size imageFlow diagram of this study.Clinical information was obtained from the electronic medical record systems of each center, clinical characteristics including the participants’ age, sex, hepatitis B virus infection, liver cirrhosis status, Child-Pugh class, and ECOG performance status. Laboratory tests included albumin, total bilirubin, prothrombin time, platelet count, lymphocyte count, gamma-glutamyl transferase, alpha-fetoprotein levels and the ALBI grade. CT imaging features were independently assessed by two radiologists (with 5 and 10 years of experience in abdominal imaging, respectively) from our team, who were blinded to the specific treatment protocols and clinical prognostic information. The evaluated imaging features included: tumor size, tumor number, extrahepatic metastasis, vascular invasion and characteristic imaging features of HCC, including arterial phase hyperenhancement (APHE), washout, and enhancing capsule.Treatment regimen and follow-upAll patients received TACE plus immunotherapy (PD-1 or PD-L1) and molecularly targeted therapy (anti-VEGF or TKIs). Patients received either standardized conventional TACE or drug-eluting microbead TACE16. The time interval between TACE and systemic treatment was less than 2 weeks. These treatment strategies and procedure was carried out by clinicians and interventional radiologists with more than 10 years of experience from participating centers. Detailed treatment procedures can be found in Supplementary Appendix S1, S2. In this study, follow-up data were obtained from outpatient and inpatient records as well as telephone follow-ups at the institution. The primary endpoint of the study was OS and the prediction models were built based on it. OS was defined as the time from the initial treatment of uHCC until either mortality or the latest follow-up occurrence (December 31, 2024).CT image acquisitionThe pretreatment DICOM format CECT images were acquired from the picture archiving and communication system (PACS). All images were obtained from multiple CT scanners (Appendix S3). The scan parameters were as follows: slice thickness, 5 mm; tube voltage, 120 kV; field of review (FOV), 40 cm×40 cm; and automated varied milliampere-second settings. Patients were injected with intravenous non-ionic iodine contrast material according to their body weight at a dose of 1.5–2 mL/kg and a rate of 2.5–3.5 mL/s. Each patient underwent imaging in four phases: plain, arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) at 30 ~ 35 s, 50 ~ 70 s, and 180 ~ 300 s, respectively.Image segmentation and feature extractionThe workflow of the radiomics analysis is shown in Fig. 2. The Dr. Wise Multimodal Research Platform (https://keyan.deepwise.com, V 3.1.2; Beijing Deepwise & League of PHD Technology Co., Ltd, Beijing, China) was used for image annotation, feature extraction, selection. A radiologist with five years of radiological reading experience independently drew the region of interest (ROI), layer by layer, along the tumor’s boundary on both AP and PVP images. Then, a senior radiologist with ten years of experience (reader 2) then reviewed and revised the ROIs. To evaluate the reproducibility of ROI delineations, 30 cases were randomly chosen for ROIs delineation by the senior radiologist two weeks later.Fig. 2Full size imageWorkflow of the method. (A) Image Acquisition, Segmentation and Feature Extraction. (B) Feature Selection and Radiomics Score Construction. (C) Clinical Factors Analysis. (D) Models Construction and Validation.Before feature extraction, we performed preprocessing procedure, including image intensity normalization and image resampling using B-spline interpolation. Simultaneously, nine different filters, including wavelet and Laplacian of Gaussian filter with different parameters, were used to preprocess the original image and transform the other images. ComBat harmonization was applied to the extracted features to correct for center-related batch effects (Appendix S4 Fig. 1).Finally, 2153 radiomics features were extracted from the AP and PVP images respectively, which included 7 classes: first-order feature, shape feature, texture feature (GLCM-Gray Level Co-occurrence Matrix, GLDM-Gray Level Dependence Matrix, GLSZM-Gray Level Size Zone Matrix, GLRLM-Gray Level Run Length Matrix, NGTDM-Neighboring Gray Tone Difference Matrix). All radiomic features were extracted using PyRadiomics v3.1.0 with its default filter configurations and default parameters. This adheres to the standardized PyRadiomics workflow (https://pyradiomics.readthedocs.io/en/latest/features.html).Feature selection and radiomics score constructionFeature selection was performed using the training cohort. Initially, radiomic features were selected for further analysis based on having inter- and intraclass correlation coefficients (ICCs) that exceeded 0.8. Next, the Pearson correlation (threshold = 0.9) and variance threshold (threshold = 0.01) methods were used to remove features with high correlation or low variation.Redundant features were removed and optimal features were selected to improve model prediction accuracy. Then, the extracted radiomics features were regularized through Z- score normalization, followed by an univariate Cox regression analysis to retain features with p