MRI radiomics and 90Y PET dosimetry for predicting hepatocellular carcinoma response after radioembolization Authors: Wenyu Huang, Jie Ding, Hongzhen Yuan, Zhidong Li, Huai Li, Yinglang Zeng Published: 2026 ABSTRACT BackgroundTo evaluate treatment response in hepatocellular carcinoma (HCC) following Yttrium-90 (90Y) radioembolization by integrating pre-treatment contrast-enhanced magnetic resonance imaging (CE-MRI) radiomics with post-treatment 90Y positron emission tomography (PET) dosimetry.Materials and methodsThis retrospective study included 57 patients with pathologically confirmed HCC. All patients underwent CE-MRI within 4–6 weeks before 90Y transarterial radioembolization (TARE), followed by 90Y PET/CT within 24 h post-treatment. Tumor response was assessed at three months using the modified Response Evaluation Criteria in Solid Tumors (mRECIST). Radiomic features extracted from CE-MRI and 24 h tumor absorbed dose parameters from 90Y PET/CT were analyzed to predict treatment response, classifying patients as responders or non-responders.ResultsBased on mRECIST criteria, 32 patients were classified as responders (all partial response), while 25 were non-responders (20 stable disease, 5 progressive disease). 24-hour average absorbed dose is independent predictive factors for treatment response (OR = 1.06, p = 0.006). The combined model, incorporating clinical characteristics and MRI-derived radiomic features, achieved a sensitivity of 88% (28/32) [95% CI, 71%–96%], specificity of 68% (17/25) [95% CI, 46%–85%], and accuracy of 79% (45/57) [95% CI, 66%–89%].ConclusionThe integration of pre-treatment CE-MRI radiomics with post-treatment 90Y PET imaging provides complementary insights into tumor biology and therapeutic efficacy. Early tumor absorbed dose metrics derived from 90Y PET, together with specific CE-MRI radiomic features, show strong correlations with treatment response and may serve as valuable predictive biomarkers to guide patient stratification and optimize radioembolization strategies in HCC.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16345-z. Introduction Hepatocellular carcinoma (HCC) remains one of the leading causes of cancer-related mortality worldwide, particularly due to the limited therapeutic options available for patients with unresectable or advanced-stage disease [1]. Among the evolving treatment modalities, transarterial radioembolization (TARE) using Yttrium-90 (90Y) microspheres has emerged as a promising locoregional therapy. This technique delivers targeted internal radiation to hepatic tumors through the arterial supply and is particularly suitable for patients who are ineligible for surgical resection or for whom other local treatments are not feasible [2, 3]. Accurate and timely assessment of treatment response is critical for determining therapeutic efficacy and informing subsequent clinical management. 90Y microsphere therapy, also known as selective internal radiation therapy (SIRT), involves the intra-arterial administration of radioactive 90Y-loaded microspheres, which selectively irradiate liver tumors while sparing surrounding healthy hepatic tissue [4, 5]. This therapeutic approach has demonstrated considerable potential in managing intermediate and advanced HCC [2, 6, 7]. Compared to conventional transarterial chemoembolization (TACE), 90Y microsphere therapy may offer several advantages, including improved patient tolerance, prolonged progression-free survival, and an increased likelihood of bridging patients to curative liver transplantation [6, 7]. A prospective, multicenter, open-label, single-arm clinical trial has evaluated the safety and efficacy of 90Y resin microspheres as a first-line treatment in patients with unresectable HCC, providing encouraging results [8]. Imaging plays a pivotal role in monitoring tumor response following 90Y radioembolization. Conventional imaging modalities such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) are frequently employed to assess therapeutic outcomes [9–11]. CT and MRI are instrumental in evaluating tumor size, morphology, and the extent of post-treatment necrosis [10–12]. In particular, multiparametric MRI techniques—such as diffusion-weighted imaging (DWI) and dynamic contrast-enhanced perfusion imaging—provide valuable insights into tumor viability and vascular response [9]. 90Y single-photon emission computed tomography (SPECT) offers a practical and widely accessible approach for visualizing the intrahepatic distribution of 90Y microspheres. It enables verification of whether the microspheres are accurately localized within the tumor and whether the absorbed radiation dose reaches therapeutically effective thresholds, thereby serving as an important tool for treatment validation and refinement [13, 14]. However, given the relatively low spatial resolution of SPECT, positron emission tomography/computed tomography (PET/CT) serves as a complementary imaging modality, providing higher-resolution and more detailed assessment of 90Y microsphere biodistribution [15, 16]. In particular, 90Y PET imaging facilitates precise post-therapy quantification of microsphere deposition, which is critical for accurate dosimetric evaluation and for optimizing dose delivery in personalized treatment planning [16]. Radiomics has emerged as a powerful tool in oncologic imaging, enabling the extraction of high-dimensional quantitative features from medical images to predict clinical outcomes in HCC patients. These features can inform on tumor biology and are increasingly used to assess treatment response, predict recurrence risk, and estimate overall survival [11]. Studies have shown that increasing the delivered radiation dose to the tumor during radioembolization correlates positively with improved imaging responses [17]. Voxel-based dosimetric analyses and dose-volume histogram (DVH) studies have further demonstrated that localized dosimetry can predict metabolic and morphological tumor responses in HCC patients treated with 90Y resin microspheres [18]. This study aims to evaluate tumor response to 90Y radioembolization in patients with HCC by employing a multimodal radiomics framework that integrates contrast-enhanced MRI and 90Y PET imaging. The primary objective is to investigate the correlation between imaging-derived radiomic features and clinical treatment response, thereby enhancing the precision of response assessment and contributing to the development of personalized treatment strategies in the management of HCC. Materials and methods Patients From February 2023 to February 2025, Xiamen Humanity Hospital conducted individualized treatment planning for 90Y microsphere radioembolization in 57 patients diagnosed with primary HCC. Each patient underwent pre-treatment simulation using 99mTc- macroaggregated albumin (MAA) SPECT/CT imaging first, followed by selective internal radiation therapy with 90Y glass microspheres. Within 24 h post-treatment, 90Y PET/CT imaging were performed to evaluate the in vivo biodistribution of microspheres. Written informed consent was obtained from all participants, and the study protocol was approved by the Ethics Committee of Xiamen Humanity Hospital (HAXM-MEC-20211118-001-01). All treatment decisions were made by a multidisciplinary tumor board. Patients were considered eligible if they were not candidates for surgical resection or liver transplantation and had at least one radiologically confirmed hepatic tumor greater than 3 cm, maintained stable liver enzyme levels within clinically acceptable limits, exhibited no contraindications to hepatic angiography, were not receiving concurrent oncologic therapies, and had no prior history of liver transplantation or hepatic resection. Exclusion criteria comprised severe hepatic or renal dysfunction, documented coagulation disorders, the presence of extrahepatic metastases, or other serious systemic illnesses that could compromise treatment safety or confound clinical outcomes. In this study, the China Liver Cancer Staging (CNLC) system was used as the primary clinical staging tool because it is the standard system recommended by the Chinese national guidelines for HCC and aligns closely with the clinical practice in our cohort [19]. Concerning tumor status, each Barcelona Clinic of Liver Cancer (BCLC) stage (0/A, B, and C) is split into two corresponding CNLC substages — specifically, CNLC Ia/Ib for BCLC 0/A, IIa/IIb for BCLC B, and IIIa/IIIb for BCLC C. CNLC stage IV matches BCLC stage D. Imaging acquisition All patients underwent treatment with 90Y microspheres. Prior to administration, the tumor-feeding artery was identified by angiography, after which the microspheres were delivered via selective catheterization of the targeted vessel. The prescribed 90Y dose was individualized according to tumor burden, liver function, and other relevant clinical parameters. PET/CT imaging was performed using a GE Discovery MI scanner (GE Healthcare, Waukesha, WI, USA). Image reconstruction employed the Ultra-TOF technique with a slice thickness of 3.27 mm and a matrix size of 256 × 256, yielding a total of 83 slices. Attenuation correction was applied during reconstruction, whereas scatter correction was not performed. Following reconstruction, a 5-mm Gaussian post-filter was applied to enhance image quality and reduce noise. Contrast-enhanced magnetic resonance imaging (CE-MRI) was employed as the standard imaging modality for all patients diagnosed with HCC. The MRI protocol included the following sequences: T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and contrast-enhanced T1-weighted imaging (CE-T1WI). Contrast-enhanced images were acquired following the intravenous administration of gadolinium-based contrast agent (gadolinium sulfate). All MRI examinations were performed using a 1.5 Tesla scanner (MAGNETOM Amira, Siemens Healthineers, Germany). DWI was not included in the radiomic analysis because not all patients underwent DWI with identical b-values and acquisition parameters due to the retrospective nature of the study. HCC segmentation MRI was chosen over CT because it provides superior soft-tissue contrast for delineating tumor boundaries and detecting subtle changes in arterial phase hyperenhancement and washout, which are critical for mRECIST assessment in HCC. Moreover, MRI-based radiomics have been shown to be more reproducible than CT for liver lesions due to the absence of beam-hardening artifacts and better delineation of necrotic components. MRI data were resampled to a uniform voxel size of 1.5 × 1.5 × 1.5 mm³ to ensure consistency in spatial resolution. Lesion segmentation of HCC on CE-T1WI was performed using the Segment Anything Model for Medical Images (MedSAM) in inference mode (Python 3.10.15, CUDA 12.1, PyTorch 2.4.0) [20]. The segmentation process began with the manual placement of a 2D bounding box around the lesion of interest, followed by iterative refinement using foreground and background prompt points to improve segmentation accuracy (Fig. 1). All final segmentation results were independently reviewed and confirmed by an experienced radiologist to ensure anatomical correctness and clinical reliability. Fig. 1Semi-automatic segmentation of hepatocellular carcinoma lesions on contrast-enhanced T1-weighted images using the MedSAM algorithm Semi-automatic segmentation of hepatocellular carcinoma lesions on contrast-enhanced T1-weighted images using the MedSAM algorithm Lesion delineation on 90Y PET/CT was performed by identifying areas of radiotracer uptake using a standardized thresholding method implemented in the 3D Slicer software (Windows version 5.4.3) and modified by manual with considering the lesion region on MRI. The lesion volume was calculated by the formula: Lesion volume = VOI_voxel_counts * voxel_volume_PET. Absorbed dose (Gy) was calculated voxel wise using the local deposition model (MIRD S16), assuming all 90Y β energy is deposited within the same voxel. The formula is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D\left(\mathrm{Gy}\right)={C}_{act}\left(\mathrm{Bq/mL}\right)\times\:\frac{49.67\:(\mathrm{Gy}\cdot\:\mathrm{kg}/\mathrm{GBq})}{\rho\:\left(\mathrm{kg/L}\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{act}$$\end{document} is the activity concentration from 90Y PET, 49.67 is the dose constant for 90Y, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:=1.04$$\end{document} kg/L for liver tissue. Dosimetry was performed with MIM SurePlan™ Liver Y90 (version 7.0). Acquisition radiomics feature and modeling Radiomic features were extracted from CE-T1WI images using the PyRadiomics (Windows version 3.1.0) [21]. The extracted features comprised a comprehensive set of descriptors, encompassing first-order statistics, texture features (e.g., gray-level co-occurrence matrix, gray-level run length matrix), and shape-based metrics. To address the issue of high dimensionality and potential overfitting, a three-step feature selection process was employed after expanding the cohort to 57 patients. First, features with poor stability (intraclass correlation coefficient [ICC] < 0.75) were excluded. Second, a correlation analysis removed redundant features with pairwise absolute correlation |r| > 0.9. Finally, least absolute shrinkage and selection operator (LASSO) regression with L1 regularization was applied to retain only the most relevant features with non-zero coefficients. Predictive modeling was performed using six machine learning algorithms: Random Forest, Extreme Gradient Boosting (XGBoost), Logistic Regression, Support Vector Machine (SVM), Naive Bayes, and K‑Nearest Neighbors (KNN). Each model was trained and validated using stratified 5‑fold cross‑validation as well as leave‑one‑out cross‑validation (LOOCV) to ensure robust and unbiased performance evaluation (Fig. 2). The predictive performance of each algorithm was assessed based on multiple metrics, including the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and Cohen’s kappa coefficient. The optimal model was selected based on its accuracy and kappa values, as these metrics best reflect overall classification performance and predictive agreement, respectively. Finally, the output of the best-performing model was visualized to enhance interpretability and support its potential application in clinical decision-making. Fig. 2Workflow for tumor segmentation, radiomic feature extraction and selection, and predictive model development for hepatocellular carcinoma treatment response Workflow for tumor segmentation, radiomic feature extraction and selection, and predictive model development for hepatocellular carcinoma treatment response Treatment response evaluation Treatment response was assessed three months post‑therapy using the modified Response Evaluation Criteria in Solid Tumors (mRECIST). We chose mRECIST as the primary response criterion because it is widely accepted in clinical trials and regulatory settings for locoregional therapy of HCC, and it directly quantifies changes in viable tumor based on arterial phase enhancement. Response was determined by comparing the relative change in the longest diameter of target lesions on axial contrast‑enhanced CT or MRI. According to mRECIST guidelines, patients were classified into two categories: responders and non-responders. The responder group included patients who met the mRECIST for either complete or partial response. Complete Response (CR) was defined as the disappearance of all intratumoral arterial enhancement in the target lesions, while Partial Response (PR) was defined as at least a 30% reduction in the sum of the diameters of viable (i.e., enhancing) target lesions. The non-responder group comprised patients classified as having Stable Disease (SD) or Progressive Disease (PD). Stable Disease was defined as insufficient tumor shrinkage to qualify for PR and insufficient tumor growth to qualify for PD. Progressive Disease was defined as at least a 20% increase in the sum of the diameters of viable (enhancing) target lesions. Statistical analysis Statistical analysis and data visualization were performed using R project (Windows version 4.2.2). Continuous variables were reported as mean ± standard deviation, while categorical variables were summarized as counts and corresponding percentages. Group comparisons were conducted using the Student’s t-test for normally distributed continuous variables and the Wilcoxon rank-sum test for non-normally distributed data. A two-sided p-value < 0.05 was considered statistically significant. For clinical feature analysis, variables with p-values < 0.1 in univariate testing were further included in the multivariate analysis. Results Patients’ characteristics A total of 57 patients with HCC who underwent 90Y radioembolization were included in this study. Pre‑treatment 99mTc‑MAA SPECT/CT and post‑treatment 90Y PET distributions were visually concordant in all 57 patients. The mean age was 61 ± 13 years, and the cohort comprised 4 females and 14 males (Table 1). The mean lesion volume was 315.15 ± 216.38 cm3, and the tumor’s 24 h average absorbed dose of 90Y was 58.55 ± 17.90 Gy. Based on mRECIST criteria, 32 patients (56%) were classified as responders, all of whom achieved partial response (PR), with no cases of complete response (CR). The remaining 25 patients (44%) were categorized as non-responders, including 20 patients (35%) with stable disease (SD) and 5 patient (9%) with progressive disease (PD). Table 1Patients basic characteristicsVariablesvaluesNumber of patients57Sex (Male/Female)Male41 (72%)Female16 (28%)Age (year)61 ± 13Aim for 90Y microsphere therapyRadiation segmentectomy3 (5%)Bridge therapy1 (2%)Downstaging therapy53 (93%c)China Liver Cancer StagingI17 (30%)II5 (9%)III35 (61%)Barcelona Clinic of Liver CancerA17 (30%)B5 (9%)C35 (61%)AJCC 8th Edition TNM ClassificationI1 (2%)II20 (35%)III36 (63%)Lesion typeSolitary lesion42 (74%)Multinodular lesions15 (26%)Macrovascular invasionYes8 (14%)No49 (86%)Lesion size (cm3)315.15 ± 216.38Administered activity (GBq)2.67 ± 1.2424 h Average absorbed dose (Gy)58.55 ± 17.90Treatment responseResponders (CR/PR)32 (0/32) [56%]Non-responders (SD/PD)25 (20/5) [44%] Table 1Patients basic characteristicsVariablesvaluesNumber of patients57Sex (Male/Female)Male41 (72%)Female16 (28%)Age (year)61 ± 13Aim for 90Y microsphere therapyRadiation segmentectomy3 (5%)Bridge therapy1 (2%)Downstaging therapy53 (93%c)China Liver Cancer StagingI17 (30%)II5 (9%)III35 (61%)Barcelona Clinic of Liver CancerA17 (30%)B5 (9%)C35 (61%)AJCC 8th Edition TNM ClassificationI1 (2%)II20 (35%)III36 (63%)Lesion typeSolitary lesion42 (74%)Multinodular lesions15 (26%)Macrovascular invasionYes8 (14%)No49 (86%)Lesion size (cm3)315.15 ± 216.38Administered activity (GBq)2.67 ± 1.2424 h Average absorbed dose (Gy)58.55 ± 17.90Treatment responseResponders (CR/PR)32 (0/32) [56%]Non-responders (SD/PD)25 (20/5) [44%] Patients basic characteristics Univariate analysis revealed no significant differences in sex distribution between responders and non-responders (10/22 vs. 6/10, p = 0.76). Similarly, age did not differ significantly between the two groups (61 ± 13 vs. 60 ± 13, p = 0.78). Liver cancer staging ((I/II/III, 13/2/17 vs. 4/3/18, p = 0.09), lesion type (solitary/multinodular, 25/7 vs. 17/8, p = 0.58), macrovascular invasion (Yes/No, 3/29 vs. 5/20, p = 0.42), administered activity (2.79 ± 1.27 vs. 2.53 ± 1.20, p = 0.42), and Lesion volume (340.62 ± 220.20 vs. 282.55 ± 211.30, p = 0.18) also showed no significant difference. However, the 24 h mean absorbed dose was significantly higher in responders compared with non-responders (64.24 ± 15.25 vs. 51.27 ± 18.68, p = 0.002) (Table 2). In multivariate analysis, the 24 h mean absorbed dose (OR = 1.06, p = 0.006) remained statistically significant predictors of treatment response (Table 2). Using the Youden index, the optimal cut-off value for predicting response was 61.65 Gy. Table 2Univariable and multivariable analysis of clinical characteristicsParametersRespondersNon-respondersUnivariate analysis (P)ORMultivariate analysis (P)Sex (Female/Male)10/226/100.760.390.21Age (year)61 ± 1360 ± 130.781.020.51China Liver Cancer Staging (I/II/III)13/2/174/3/180.090.430.06Lesion type (Solitary/Multinodular)25/717/80.580.370.17Macrovascular invasion (Yes/No)3/295/200.421.110.91Administered activity (GBq)2.79 ± 1.272.53 ± 1.200.421.210.69Lesion volume (cm3)340.62 ± 220.20282.55 ± 211.300.181.000.9524 h Average absorbed dose (Gy)64.24 ± 15.2551.27 ± 18.680.0021.060.006*P < 0.05* Table 2Univariable and multivariable analysis of clinical characteristicsParametersRespondersNon-respondersUnivariate analysis (P)ORMultivariate analysis (P)Sex (Female/Male)10/226/100.760.390.21Age (year)61 ± 1360 ± 130.781.020.51China Liver Cancer Staging (I/II/III)13/2/174/3/180.090.430.06Lesion type (Solitary/Multinodular)25/717/80.580.370.17Macrovascular invasion (Yes/No)3/295/200.421.110.91Administered activity (GBq)2.79 ± 1.272.53 ± 1.200.421.210.69Lesion volume (cm3)340.62 ± 220.20282.55 ± 211.300.181.000.9524 h Average absorbed dose (Gy)64.24 ± 15.2551.27 ± 18.680.0021.060.006*P < 0.05* Univariable and multivariable analysis of clinical characteristics P < 0.05* Clinical model performance A clinical prediction model was constructed using logistic regression, incorporating age, sex, lesion volume, and mean tumor absorbed dose as predictor variables. The model achieved a sensitivity of 69% (22/32) [95% CI, 50%–84%], a specificity of 64% (16/25) [95% CI, 43%–82%], and an overall accuracy of 67% (38/57) [95% CI, 53%–79%]. The area under the receiver operating characteristic curve (AUC) was 0.73 [95% CI, 0.59–0.86], and the Cohen’s kappa coefficient was 0.33, indicating moderate agreement between predicted and observed treatment responses. Feature selection and performance of radiomics models After dimensionality reduction using LASSO regression, 10 MRI-derived radiomic features were retained for model development. Among the evaluated machine learning algorithms, the Random Forest model demonstrated the highest predictive performance, achieving a sensitivity of 91%(29/32)[75%-98%], specificity of 64%(16/25)[43%-82%], overall accuracy of 79%(45/57)[66%-89%], an AUC of 0.73[0.59–0.88], and a Cohen’s kappa coefficient of 0.56, reflecting moderate agreement. Feature importance and overall feature contribution are shown in Fig. 3, which highlights the key predictors of treatment response. Gray-level and shape features exerted the greatest influence on model output, and these features jointly drove the prediction outcome. The KNN model also showed comparable performance, with an accuracy of 77%(44/57)[64%-87%] and a kappa value of 0.52 (Table 3). Fig. 3Interpretation of key predictors for treatment response in the optimal random forest model. Relative importance of the top features contributing to the prediction A. SHAP value plots illustrating the direction and magnitude of each feature’s influence on the predicted outcome B Interpretation of key predictors for treatment response in the optimal random forest model. Relative importance of the top features contributing to the prediction A. SHAP value plots illustrating the direction and magnitude of each feature’s influence on the predicted outcome B Table 3Performance of the prediction model based on the MRI-radiomicsSensitivity [95% CI]Specificity [95% CI]Accuracy [95% CI]AUC [95% CI]KappaRandom Forest91%(29/32)[75%-98%]64%(16/25)[43%-82%]79%(45/57)[66%-89%]0.73[0.59–0.88]0.56XGBoost78%(25/32)[60%-91%]56%(14/25)[35%-76%]68%(39/57)[55%-80%]0.73[0.59–0.87]0.35Logistic Regression75%(24/32)[57%-89%]60%(15/25)[39%-79%]68%(39/57)[55%-80%]0.77[0.64–0.90]0.35Naive Bayes81%(26/32)[64%-93%]56%(14/25)[35%-76%]70%(40/57)[57%-82%]0.80[0.69–0.92]0.38SVM84%(27/32)[67%-95%]48%(12/25)[28%-69%]68%(39/57)[55%-80%]0.78[0.65–0.90]0.34KNN94%(30/32)[79%-99%]56%(14/25)[35%-76%]77%(44/57)[64%-87%]0.78[0.66–0.90]0.52 Table 3Performance of the prediction model based on the MRI-radiomicsSensitivity [95% CI]Specificity [95% CI]Accuracy [95% CI]AUC [95% CI]KappaRandom Forest91%(29/32)[75%-98%]64%(16/25)[43%-82%]79%(45/57)[66%-89%]0.73[0.59–0.88]0.56XGBoost78%(25/32)[60%-91%]56%(14/25)[35%-76%]68%(39/57)[55%-80%]0.73[0.59–0.87]0.35Logistic Regression75%(24/32)[57%-89%]60%(15/25)[39%-79%]68%(39/57)[55%-80%]0.77[0.64–0.90]0.35Naive Bayes81%(26/32)[64%-93%]56%(14/25)[35%-76%]70%(40/57)[57%-82%]0.80[0.69–0.92]0.38SVM84%(27/32)[67%-95%]48%(12/25)[28%-69%]68%(39/57)[55%-80%]0.78[0.65–0.90]0.34KNN94%(30/32)[79%-99%]56%(14/25)[35%-76%]77%(44/57)[64%-87%]0.78[0.66–0.90]0.52 Performance of the prediction model based on the MRI-radiomics Performance of combined clinical and MRI-based radiomics models When clinical features were integrated with MRI-based radiomics features, the predictive model demonstrated improved diagnostic performance (Table 4). The KNN model achieved the best results, with a sensitivity of 88%(28/32)[71%-96%], specificity of 68%(17/25)[46%-85%], overall accuracy of 79%(45/57)[66%-89%], an AUC of 0.80[0.68–0.91], and a kappa value of 0.56. The Random Forest model also performed strongly, yielding the same accuracy (79%[66%-89%]) and kappa (0.56). Table 4Performance of the prediction model based on the combined clinical feature and MRI-radiomicsSensitivity [95% CI]Specificity [95% CI]Accuracy [95% CI]AUC [95% CI]KappaRandom Forest91%(29/32)[75%-98%]64%(16/25)[43%-82%]79%(45/57)[66%-89%]0.79[0.66–0.92]0.56XGBoost78%(25/32)[60%-91%]56%(14/25)[35%-76%]68%(39/57)[55%-80%]0.76[0.63–0.89]0.35Logistic Regression75%(24/32)[57%-89%]72%(18/25)[51%-88%]74%(42/57)[60%-84%]0.81[0.69–0.93]0.47Naive Bayes81%(26/32)[64%-93%]68%(17/25)[46%-85%]75%(43/57)[62%-86%]0.82[0.72–0.93]0.50SVM91%(29/32)[75%-98%]52%(13/25)[31%-72%]74%(42/57)[60%-84%]0.77[0.64–0.90]0.44KNN88%(28/32)[71%-96%]68%(17/25)[46%-85%]79%(45/57)[66%-89%]0.80[0.68–0.91]0.56 Table 4Performance of the prediction model based on the combined clinical feature and MRI-radiomicsSensitivity [95% CI]Specificity [95% CI]Accuracy [95% CI]AUC [95% CI]KappaRandom Forest91%(29/32)[75%-98%]64%(16/25)[43%-82%]79%(45/57)[66%-89%]0.79[0.66–0.92]0.56XGBoost78%(25/32)[60%-91%]56%(14/25)[35%-76%]68%(39/57)[55%-80%]0.76[0.63–0.89]0.35Logistic Regression75%(24/32)[57%-89%]72%(18/25)[51%-88%]74%(42/57)[60%-84%]0.81[0.69–0.93]0.47Naive Bayes81%(26/32)[64%-93%]68%(17/25)[46%-85%]75%(43/57)[62%-86%]0.82[0.72–0.93]0.50SVM91%(29/32)[75%-98%]52%(13/25)[31%-72%]74%(42/57)[60%-84%]0.77[0.64–0.90]0.44KNN88%(28/32)[71%-96%]68%(17/25)[46%-85%]79%(45/57)[66%-89%]0.80[0.68–0.91]0.56 Performance of the prediction model based on the combined clinical feature and MRI-radiomics Based on the logistic regression model, the ROC curves showed that the combined model achieved superior performance (accuracy, 74%) compared with the clinical model (accuracy, 67%) and the radiomics-only model (accuracy, 68%) (Fig. 4). Among the clinical variables, the 24 h average absorbed dose emerged as a relatively important predictor. Based on these combined clinical and radiomic predictors, a personalized nomogram was constructed (Fig. 5), with representative cases illustrated in Figure A1. Fig. 4The performance of clinical model, radiomics model, and combined model in predicting treatment response The performance of clinical model, radiomics model, and combined model in predicting treatment response Fig. 5Nomogram integrating clinical and radiomic features to estimate the probability of treatment response following 90Y TARE in hepatocellular carcinoma Nomogram integrating clinical and radiomic features to estimate the probability of treatment response following 90Y TARE in hepatocellular carcinoma Discussion This study demonstrates that integrating pre-treatment CE-MRI radiomic features with post-treatment 90Y PET-derived dosimetric parameters achieves robust predictive performance for treatment response in patients with HCC undergoing 90Y TARE. In our cohort, the combined model yielded a sensitivity of 88%, specificity of 68%, and overall accuracy of 79%, highlighting the potential of multimodal imaging biomarkers to improve patient selection and treatment planning. 90Y PET enables quantification of tumor absorbed dose distribution, providing critical information for evaluating treatment response [22]. In our cohort, the mean absorbed dose was significantly higher in responders than in non-responders, consistent with previous 90Y PET dosimetry studies [23]. Previous reports have also identified absorbed dose as an independent predictor of treatment response. However, while some studies have cited thresholds around 40 Gy as strong determinants, our findings suggest a higher optimal cut-off of 61.65 Gy [24]. It should be noted that this analysis was based on the present cohort; thus, the identified cut‑off is exploratory and not validated, and its stability requires confirmation in larger cohorts. Importantly, when average absorbed dose was integrated with pre-treatment CE-MRI radiomics features, the overall predictive accuracy was substantially improved, underscoring the added value of a multimodal approach. Radiomics-based prediction models have been investigated in other HCC treatment settings, such as transarterial chemoembolization and radiofrequency ablation [25–28], but their application to 90Y TARE remains limited. Importantly, previous studies often overlooked tumor absorbed dose and absorbed volume, both of which have a critical impact on therapeutic efficacy [29]. MRI-based radiomics models using both pre- and post-treatment images have reported sensitivities of approximately 91% and specificities of 78% [30]. In contrast, our study demonstrated that even a pre-treatment MRI radiomics model alone achieved good predictive accuracy, highlighting that clinically meaningful information can be extracted before therapy initiation. Moreover, our findings are consistent with growing evidence that integrating radiomics with dosiomics or dosimetry metrics can generate robust predictive models, as demonstrated in recent work using SPECT-derived radiomics [31]. From a feature interpretability perspective, this study focuses on the combined contribution of multiple radiomic and dosimetric features rather than relying on any single biomarker. Given the potential instability of feature values or feature combinations caused by variations in MRI acquisition conditions, batch effects, and other factors, emphasizing the overall contribution of radiomic and clinical features is more meaningful. The aforementioned features provide a quantitative framework for translating complex imaging patterns into measurable biomarkers of treatment response. However, their stability and reproducibility across different imaging platforms, acquisition protocols, and patient populations require further validation through larger, multi-center prospective studies. The proposed predictive framework naturally lends itself to the development of a personalized clinical nomogram, enabling individualized risk estimation in a visually intuitive format [32]. Such a tool could support risk-adapted surveillance schedules, inform decisions on adjunctive therapies, and potentially guide patient selection for 90Y TARE. While the retrospective design of this study restricts the generalizability of our findings, the observed performance metrics suggest that with prospective validation and larger cohorts, this approach could be translated into routine clinical practice. Several limitations warrant consideration. First, this was a single-center, retrospective study with a modest sample size, which may limit statistical power and increase the risk of overfitting despite our dimensionality reduction efforts. Second, the absence of an external validation cohort means that the performance estimates may be optimistic and not directly generalizable to other institutions or scanner platforms. Third, we used mRECIST at three months as the primary endpoint rather than longer-term outcomes such as progression-free survival or overall survival; thus, the clinical benefit of early response prediction remains to be established. Finally, the lack of scatter correction during 90Y PET reconstruction may have introduced a modest systematic bias in absolute dose quantification, although relative comparisons and model performance appear robust. In summary, this work provides compelling evidence that combining CE-MRI radiomics with 90Y PET dosimetry can deliver highly accurate, noninvasive prediction of treatment response to 90Y TARE in HCC. This strategy offers a practical pathway toward precision oncology in interventional radiology, with the potential to optimize patient outcomes and resource utilization. Future multicenter, prospective studies with external validation are essential to confirm the robustness and clinical utility of these findings. Supplementary Information Supplementary Material 1. Supplementary Material 1.