Voxel-based dosimetry for predicting treatment response to transarterial radioembolization in hepatocellular carcinoma: significance of intratumoral dose distribution Authors: Minseok Suh, Keon Min Kim, Jin Woo Choi, Jae Sung Lee, Jin Chul Paeng, Hyo-Cheol Kim Published: 2025 ABSTRACT AbstractPurposeWe applied a voxel-based personalized dosimetry method using multiple voxel S-values (VSVs) to assess the intratumoral dose distribution in hepatocellular carcinoma (HCC) patients treated with trans-arterial radioembolization (TARE) using glass microspheres. This study aimed to evaluate the predictive value of dose metrics for treatment response and local progression-free survival (L-PFS).MethodsNinety patients with HCC who underwent TARE between November 2015 and December 2019 were retrospectively analyzed. Post-treatment 90Y-microsphere PET/CT images were used to generate voxel-wise absorbed dose maps via convolution with CT-derived multiple VSV kernels. Tumor volumes were manually delineated on contrast-enhanced CT co-registered with dose map. Dose-volume histogram (DVH) parameters, including V-V205 (tumor volume receiving < 205 Gy) and D70, were calculated. Dose heterogeneity was assessed using the coefficient of variation (CoV) of voxel doses. The performance of dose parameters for predicting complete response (CR) and L-PFS was evaluated using logistic and Cox regression analyses.ResultsAmong 90 patients, 57 (63.3%) achieved CR. PET average dose, D70, V-V205, and CoV significantly differed according to treatment response. ROC analysis showed good predictive performance for CR with D70 (AUC 0.908), V-V205 (0.903), PET average dose (0.876), and CoV (0.870). In multivariate logistic regression, only small V-V205 (< 22.8 mL; OR 12.50, P = 0.002) remained an independent predictor of CR. For L-PFS, multivariate Cox analysis identified V-V205 < 22.8 mL (HR 0.07, P = 0.013) and CR (HR 0.29, P = 0.034) as independent prognostic factors.ConclusionVoxel-based dosimetry using multiple VSV kernels enables quantitative assessment of intratumoral dose distribution in HCC patients treated with TARE. Among voxel-level parameters, D70, V-V205, and CoV showed good performance in predicting CR, and V-V205 was the only independent predictor for both treatment response and L-PFS. These findings support the added prognostic value of voxel-based dose metrics beyond average tumor absorbed dose.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40658-025-00822-8.Keywords: Trans-arterial radioembolization, Voxel-based dosimetry, Tumor heterogeneity, Hepatocellular carcinoma Introduction Trans-arterial radioembolization (TARE) is a selective internal radiation therapy that injects a 90Y labeled microsphere directly into the tumor tissue through the hepatic artery. As a result, it is possible to deliver a higher dose to the tumor and at the same time minimize the exposure to surrounding normal parenchyma. Retrospective cohort studies and prospective clinical trials have revealed the efficacy and safety of TARE across different stages of hepatocellular carcinoma (HCC) [1–5]. However, recent studies have presented conflicting results. In two randomized controlled trials, comparing the efficacy with Sorafenib, TARE was better tolerated but failed to prove superiority in overall survival or progression-free survival [6, 7]. A similar result was reported in the meta-analysis comparing TARE to the standard of care in non-surgical HCC patients [8]. One major point of the conflicting result is that no endpoint regarding tumor-absorbed dose was considered in the studies [9, 10]. The tumor dose–response relationship has been investigated in TARE [11]. In the studies, the difference in absorbed dose between responder and non-responder was observed through pre-therapeutic dosimetry using 99mTc-macro aggregated albumin (MAA) SPECT or post-therapeutic dosimetry based on 90Y PET or SPECT, and threshold doses for the prediction of treatment response were suggested [12–15]. Furthermore, recent studies have investigated the impact of dose intensification guided by personalized dosimetry, demonstrating the feasibility of delivering a boosted radiation dose tailored to individual patients [16–18]. Building on these findings, the multicenter phase 2 DOSISPHERE-01 trial showed that personalized dose escalation significantly improves the objective response rate in patients with locally advanced HCC [19]. Another point of consideration is heterogeneous intratumoral dose distribution. HCC patients undergoing TARE often have large, necrotic, and heterogeneous tumors. These tumors usually show a non-uniform distribution of microspheres, and the average tumor absorbed dose alone may not accurately predict the therapeutic response. Furthermore, intratumoral heterogeneity is a well-recognized hallmark of HCC progression [20]. Accordingly, recent studies have evaluated intratumoral dose distribution using voxel-level analyses, including dose-volume histograms (DVH) and equivalent uniform dose derived from voxel-based dosimetry [21–24]. However, the added value of assessing intratumoral dose distribution compared to relying solely on the average tumor absorbed dose has not yet been fully clarified [23, 25–28]. We have recently established a simple voxel-based personalized dosimetry method using multiple voxel S-values (VSVs) [29]. In the pilot study, we applied this method and evaluated the efficacy for predicting response in HCC patients who underwent 90Y-Resin microsphere [30]. The average tumor absorbed dose estimated on post-treatment 90Y PET was effective in predicting prognosis, but heterogeneity indices were not. To further confirm the clinical significance of intratumoral dose distribution we applied voxel-based dosimetry for each tumor in HCC patients who underwent TARE and we evaluated the impact of dosimetry parameters to predict treatment response and local progression-free survival (L-PFS). Method Patients Patients with HCC, who underwent TARE with 90Y-glass microsphere (Therasphere®) and post-TARE 90Y-microsphere PET/CT scan between November 2015 to December 2019 in our institution, were retrospectively retrieved. Among 256 patients, we retrospectively analyzed patients who met the following inclusion and exclusion criteria. Inclusion criteria: (1) Four or fewer tumors, (2) 3 cm or larger, (3) no major vascular invasion, (4) no previous chemoembolization, ablation therapy, and systemic therapy (sorafenib) for the primary target tumor, and (5) at least three-month imaging follow-up. Patients who had (1) infiltrative HCC, (2) scheduled surgical resection after radioembolization, (3) administration of any additional locoregional or systemic therapy after TARE and before best response assessment, and (4) follow-up loss within three months were excluded from the study. This study shares a patient cohort with previously published literature, and the previous study dealt with the clinical significance of virtual tumor absorbed dose, a dose calculated by assuming that all infused microspheres were distributed only in the tumor [31]. The study was approved by the Institutional Review Board of our institution (Approved No. 2207–030-1337), and written informed consent was waived. Trans-arterial radioembolization TARE was conducted by interventional radiologists with more than 10 years of experience in vascular intervention as previously described [18]. Briefly, 99mTc-MAA was injected into the tumor-feeding artery and subsequently, a planar scan and SPECT/CT were performed for the treatment planning. Details of the SPECT/CT acquisition and reconstruction protocols are provided in the Supplementary Methods. Boosted radioembolization was performed by delivering > 150 Gy to the perfused lobe, when at least two liver segments were spared and the estimated lung dose was < 30 Gy. Therasphere® was delivered through the hepatic artery to the tumors with preferential blood flow according to a standardized protocol [32]. The pre-therapeutic tumor dose (pre-Dose) was calculated using the Medical Internal Radiation Dose (MIRD) schema within a multi-compartment model, based on 99mTc-MAA SPECT/CT imaging. Regions of interest (ROIs) for tumor, normal liver, and lungs were segmented on co-registered CT, and the activity distribution was derived from MAA uptake. PET image acquisition 90Y-microsphere PET/CT image was acquired as previously described, using a large field-of-view PET/CT scanner (Biograph mCT64, Siemens Healthineers, Germany; axial field-of-view, 216 mm) immediately after TARE [30]. PET image was confined to a single-bed position covering from the lower chest to the upper abdomen and acquired for 10 min in 3D mode. The images were reconstructed on 200 × 200 matrices with a voxel size of 4.07 × 4.07 × 5 mm3 using a 3D ordered subsets expectation maximization (OSEM) algorithm (2 iterations, 21 subsets), incorporating point-spread function (PSF) modeling and time-of-flight (TOF) estimation. A Gaussian post-reconstruction filter with a full-width at half maximum (FWHM) of 5 mm was applied. CT was acquired with a slice thickness of 5 mm, 1.2 pitch, 120 kVp, and 35 mAs for attenuation correction. Voxel-based dosimetry The process for the acquisition of dosimetry parameters is illustrated in Fig. 1. Process of voxel-based dosimetry An absorbed dose map was generated as previously described with some modification [30]. Briefly, we used a VSV (voxelized S-value) kernel (Gy/MBq·s) convolution approach, where the VSV is defined as the mean absorbed dose in a target voxel per radioactive decay in a source voxel. The voxel dose (D) was calculated by convolving the VSV with the cumulated activity (Ã) using the following equation: Firstly, we used six VSV kernels for 90Y generated by GATE simulation, representing different media with various densities (air, lung, 0.9 g/cm3, water, 1.1 g/cm3, and bone). Each kernel had a matrix size of 51 × 51 × 45 and a voxel size of 4.07 × 4.07 × 5.00 mm3, which was identical to the PET image voxel size; thus, no resampling of PET images was performed. The details of the elemental composition, density, and matrix size for each VSV kernel are provided in Supplementary Table 1. Secondly, accumulated activity maps were generated from the resized PET image by using the physical half-life. Lastly, the 3D cumulated activity map (MBq·s) was convoluted with the 3D VSV (Gy/MBq·s) to yield an absorbed dose map (Gy). The absorbed dose map was fused with the most recent contrast-enhanced CT image obtained prior to TARE (within 14 days for all cases, median interval: 6 days), using a commercial deformable registration package (MIM EncoreTM, MIM Software Inc., Cleveland, OH). Firstly, manually adjusted rigid registration is performed between CT of PET/CT and contrast-enhanced CT. Then, the deformable registration and adaptive contouring module in MIM software deform the contrast-enhanced CT to match the CT of PET/CT. Finally, the absorbed dose map was fused with a deformed contrast-enhanced CT image. The ROIs for a tumor were drawn manually based on enhanced CT images on every image slice and interpolated to acquire a single volume of interest (VOI) for a tumor. PET average dose was measured from the tumor VOI. DVH was generated from the tumor VOI and V-V205 (ml) was calculated. V-V205 was defined as the tumor volume receiving a dose lower than 205 Gy. Currently, there is no standard for reporting DVH for radioembolization. Therefore, the threshold of 205 Gy was selected based on prior research that distinguished complete and incomplete responses to treatment [19]. We also determined D70, the minimum dose received by 70% of the tumor volume, a widely used dose‑volume benchmark in radiation therapy [24]. Intratumoral heterogeneity of dose distribution was assessed using the coefficient of variation (CoV), defined as the ratio between the standard deviation and mean tumor dose. Endpoints and assessments Follow-up data and clinical information, including Barcelona clinic liver cancer (BCLC) class and Child–Pugh class, for patients’ outcomes, were retrieved from the health information system. Patients were followed regularly (or when unusual symptoms or signs appeared) by radiological examinations for an average of 23.8 months ± 13.6 (median 21.9 months; range = 3.2–53.4 months). Two interventional radiologists, with more than 10 years of experience, evaluated the tumor response and presence of local progressive disease. In the case of multinodular HCC, the largest tumor was evaluated. The primary endpoint was the prediction of the best overall response. The response to therapy was evaluated based on the modified RECIST criteria for HCC [33]. Briefly, complete loss of tumoral enhancement of the primary tumor in the arterial phase was defined as complete response (CR), a 30% decrease in the diameter of viable target lesion was defined as partial response (PR), an increase of at least 20% in the diameter of viable target lesion was recorded as progressive disease (PD), and patients who did not belong to any other category were classified as stable disease (SD). The secondary endpoint was L-PFS, defined as the time between the TARE and the confirmed progression. Local progressive disease was defined as the appearance of a new lesion within the target tumor or the presence of PD finding of the target lesion. If surgical resection or chemoembolization was performed for the primary target tumor, local progression-free survival was censored on the day of the procedure. Statistical analysis All Statistical analyses were performed using PRISM version 5.0 (GraphPad Software, San Diego, CA, USA) and MedCalc statistical packages version 14.8 (MedCalc Statistical Software, Mariakerke, Belgium). The correlation between dosimetry parameters was determined using the Pearson correlation coefficient. Groupwise comparison of quantitative measurement was performed using the Kruskal–Wallis test and subsequent post hoc analyses. Receiver-operating characteristic (ROC) curves were generated and the area under the ROC curve (AUC) was calculated. The optimal cutoff values were acquired from the ROC curves for dosimetry parameters to predict the best overall response. To evaluate predictors of complete response (CR), univariate and multivariable logistic regression analyses were performed. Due to potential multicollinearity among PET-based dosimetric variables, all candidate variables were entered into a forward stepwise logistic regression using Akaike Information Criterion (AIC) for model selection. The final model was selected based on the lowest AIC value, identifying the most parsimonious set of independent predictors of CR. Cox proportional hazard model was used for univariate and multivariate analysis to evaluate the prognostic impact of clinical characteristics and dosimetry parameters in terms of L-PFS. Results A total of 90 patients with HCC were evaluated in the study. Patient characteristics are summarized in Table 1. Briefly, most of the patients were classified as low risk in the classification system, 62.2% as BCLC class A and 98.8% as Child–Pugh class A. The tumor was mostly single nodular (67.8%) and limited to one lobe (63%). The median tumor volume was 98.0 mL (8.7–1574.4). The patients had a mean age of 64.3 (± 12.2) years. Table 1Patients characteristicsCharacteristicsMean ± SDAge64.3 ± 12.2No. of patients (%)GenderMale75 (83.3)Female15 (16.7)BCLC classA56 (62.2)B26 (28.9)C8 (8.9)Child Pugh ClassA576 (84.4)A613 (14.4)B81 (1.1)Number of lesionsSingle nodular61 (67.8)Multinodular29 (32.2)Lobar involvementUnilobar63 (70.0)Bilobar27 (30.0)Median (IQR), range (ml)Tumor volume98.0 (33.4–276.5), 8.7–1574.4SD, standard deviation; BCLC, Barcelona clinic liver cancer; IQR, interquartile range Patients characteristics SD, standard deviation; BCLC, Barcelona clinic liver cancer; IQR, interquartile range Dosimetry parameters The mean values of pre-Dose, PET average dose, CoV, D70, and V-V205 were 260.4 Gy (range 84 to 700 Gy), 490.2 Gy (range 16.3 to 1416.5 Gy), 0.6 (range 0.2 to 3.3), 422.3 Gy (range 15.2 to 1238.9 Gy), and 158.6 ml (range 0 to 1476.3 ml), respectively. We assessed correlations among dosimetry parameters to explore their interrelationships. PET average dose showed moderate positive correlation with pre-Dose (r = 0.55, P < 0.001) and D70 (r = 0.63, P < 0.001), and negative correlation with V-V205 (r = –0.42, P < 0.001) and CoV (r = –0.37, P < 0.001). V-V205 also showed strong inverse correlation with D70 (r = –0.75, P < 0.001), supporting their complementary roles. These relationships are summarized in a correlation heatmap (Supplementary Fig. 1). Best overall response Based on the best overall response, 57 (63.3%) patients had CR, 22 (24.4%) had PR, and 11 (12.2%) had SD. Pre-Dose (P = 0.001), PET average dose (P < 0.001), CoV (P < 0.001), D70 (P < 0.001), and V-V205 (P < 0.001) were significantly different according to the best overall response (Fig. 2A-D). Post hoc analysis revealed that PET average dose and V-V205 were significantly different between the CR, PR, and SD groups. In contrast, D70, CoV, and pre-Dose showed no significant difference between the PR and SD groups. In ROC analysis for predicting CR, D70 had the greatest AUC value of 0.908 (95% CI: 0.834–0.961), followed by V-V205 (0.903, 95% CI: 0.827–0.963), PET average dose (0.876, 95% CI: 0.790–0.936), and CoV (0.870, 95% CI: 0.783–0.932) (Fig. 2E). Pre-Dose had the lowest AUC of 0.711 (95% CI: 0.605–0.801). Pre-Dose had a significantly lower AUC for predicting CR compared to all PET-based metrics (PET average dose, D70, CoV, and V-V205), while there were no significant differences among the PET-based metrics. The optimal cutoff values for predicting CR were 286 Gy for pre-Dose, 471.2 Gy for PET average dose, 0.6 for CoV, 231.5 Gy for D70, and 22.8 ml for V-V205. Differences in A Pre-dose, B PET average dose, C V-V205, D D70, and E coefficient of variance (CoV) according to the best overall response. The mean with standard deviation is plotted as error bars. Kruskal–Wallis test with post hoc pairwise comparisons were used for statistical analysis. Upper horizontal segments indicate statistically significant differences between groups (p < 0.05). (F) Receiver operating characteristic (ROC) curve analysis for predicting complete response. Asterisks indicate statistically significant differences (p < 0.05) In univariate analyses, small tumor size (< 7 cm), unilobar involvement, Child–Pugh score of 5, high pre-Dose (> 286 Gy), high PET average dose (> 471.2 Gy), small V-V205 (< 22.8 ml), high D70 (> 231.5 Gy), and small CoV (< 0.6) were significant predictors for CR (Table 2). In the multivariate analysis based on forward stepwise selection, only small V-V205 (OR 12.50, P = 0.002) remained a significant predictor for CR. Table 2Predictors of best overall response: univariable and multivariable logistic regression analysisUnivariable analysisMultivariable analysisOR (95% CI)POR (95% CI)PTumor size (< 7 cm)20.30 (6.11–67.41) < 0.0015.1 (1.02–25.8)0.047Unilobar involvement10.99 (3.83–31.56) < 0.0015.0 (1.21–20.5)0.026Child Pugh Class (A5)15.71 (3.24–76.23) < 0.001Pre-Dose (> 286 Gy)9.66 (2.64–35.27) < 0.001PET average dose (> 471.2 Gy)31.00 (6.70–143.50) < 0.001D70 (> 231.5 Gy)24.6 (7.30–82.50) < 0.001V-V205 (< 22.8 ml)47.00 (11.95–184.08) < 0.00112.5 (2.54–61.40)0.002CoV (< 0.6)22.32 (7.27–68.57) < 0.001D70, minimum dose received by 70% of the tumor volume; V-V205, tumor volume receiving a dose lower than 205 Gy; CoV, coefficient of variation Predictors of best overall response: univariable and multivariable logistic regression analysis D70, minimum dose received by 70% of the tumor volume; V-V205, tumor volume receiving a dose lower than 205 Gy; CoV, coefficient of variation Subgroup analysis was performed in patients with a low PET average dose (< 471.2 Gy). In the comparative analysis between patients who achieved CR and those with PR or SD, the CR group exhibited significantly lower V-V205 (P < 0.001) and coefficient of variation (CoV) (P < 0.001), as well as significantly higher D70 (P < 0.001) and PET average dose (P = 0.031) (Fig. 3A-C). ROC analysis revealed that D70 (0.791, 95% CI: 0.656–0.908), V-V205 (0.798, 95% CI: 0.661–0.919), and CoV (0.837, 95% CI: 0.707–0.946) demonstrated good performance in predicting CR (Fig. 3D). In pairwise comparisons, D70 demonstrated a significantly greater AUC than PET average dose (P < 0.001), whereas V-V205 and CoV also showed higher AUCs than PET average dose, but these differences were not statistically significant. Subgroup analysis of patients with a low PET average dose (< 471.2 Gy). Difference of A PET average dose, B D70, C V-V205, D coefficient of variance (CoV), between complete response and stable disease or progressive disease. Median with the standard deviation as an error bar was plotted on the chart. Mann–Whitney test was used for statistical comparison. (E) Receiver operating characteristic (ROC) curve analysis for predicting complete response Local progression-free survival During follow-up, local progression was observed in 19 of 90 patients (21.1%). In the univariate Cox regression analysis, several variables were significantly associated with improved L-PFS, including CR, small tumor size (< 7 cm), Child–Pugh score of 5, high PET average dose (> 471.2 Gy), small V-V205 (< 22.8 mL), high D70 (> 231.5 Gy), and small CoV (< 0.6) (Table 3). Among these, multivariable analysis using forward stepwise selection identified two independent prognostic factors: best overall response of CR (HR 0.29, 95% CI: 0.09–0.91, p = 0.034) and V-V205 < 22.8 mL (HR 0.07, 95% CI: 0.01–0.56, p = 0.013). The Kaplan–Meier analysis indicated a significant difference in survival outcomes (p < 0.001), with V-V205 < 22.8 ml having a lower risk of progression (Fig. 4). Table 3Predictors of local progression-free survival: univariable and multivariable Cox proportional hazards analysisUnivariable analysisMultivariable analysisHR (95% CI)PHR (95% CI)PBest response (CR)0.10 (0.04–0.29) < 0.0010.29 (0.09–0.91)0.034Tumor size (< 7 cm)0.13 (0.04–0.44)0.001Unilobar involvement0.303Child Pugh Class (A5)0.30 (0.11–0.84)0.022Pre-Dose (> 286 Gy)0.944PET average dose (> 471.2 Gy)0.06 (0.01–0.46)0.007D70 (> 231.5 Gy)0.08 (0.02–0.35) < 0.001V-V205 (< 22.8 ml)0.03 (0.01–0.26)0.0010.07 (0.01–0.56)0.013CoV (< 0.6)0.16 (0.06–0.46) < 0.001CR, complete response; D70, minimum dose received by 70% of the tumor volume; V-V205, tumor volume receiving a dose lower than 205 Gy; CoV, coefficient of variation Predictors of local progression-free survival: univariable and multivariable Cox proportional hazards analysis CR, complete response; D70, minimum dose received by 70% of the tumor volume; V-V205, tumor volume receiving a dose lower than 205 Gy; CoV, coefficient of variation Kaplan–Meier survival analysis of local progression-free survival stratified by V-V205, with thresholds at < 22.8 ml and ≥ 22.8 ml Discussion Theoretically, assessment of intratumoral dose distribution has advantages over average tumor absorbed dose, which does not consider heterogeneous microsphere distribution. Studies have been conducted to predict the treatment effect by applying dose metrics that reflect heterogeneous distribution, but to date, no clear advantage over the average tumor dose was observed [21–24, 26–28]. In the present study, the intratumoral dose distribution was assessed based on voxel-based dosimetry using the dose metrics of CoV, D70, and V-V205, which were significant predictors for treatment response in HCC patients undergoing TARE. Particularly in the subgroup of tumors with a low PET average dose, V-V205, D70, and CoV showed good performance in predicting response. V-V205 was the only remaining significant prognostic factor for patients’ local progression. However, these voxel-based dose metrics showed high inter-correlation in our cohort, indicating that their individual prognostic significance should be interpreted with caution due to potential collinearity. Overall, we have demonstrated the added value of dose metrics based on voxel-based dosimetry. We have defined a novel dose metric, V-V205, based on the dose–volume histogram (DVH), which summarizes 3-dimensional intratumoral dose distributions in a graphical 2-dimensional format and is widely used in treatment planning for radiation therapy [34]. Kao et al. first reported dose–response results for resin microspheres using the DVH‑derived D70 metric, with a threshold of D70 > 100 Gy to predict complete response [24]. In our glass‑microsphere cohort, we found that D70 > 231.5 Gy predicted complete response with high accuracy. Given that D70 reflects the dose coverage to the majority of the tumor volume, the complementary concept, quantifying the subtherapeutic portion, may also inform treatment efficacy. Building on this concept, V-V205 was defined as the absolute tumor volume receiving less than 205 Gy, under the hypothesis that a greater subtherapeutic dose volume would correlate with lower local response and a higher risk of recurrence [35]. As V-V205 quantifies the absolute volume of tumor receiving subtherapeutic radiation, it may serve as a surrogate for intratumoral under-treatment and provide complementary information to other dose metrics in predicting treatment response. V‑V205 also correlates with total tumor volume (r = 0.865, P < 0.001), reflecting the fact that larger lesions naturally contain more under‑dosed voxels; thus, tumor size itself may in part drive the observed association with best overall response. CoV is a first-order texture parameter reflecting the statistical heterogeneity of the region of interest [36]. CoV has been widely used to predict therapy response and prognosis in oncologic PET/CT [36–40]. CoV is an intuitive and accessible parameter compared to other 2nd or higher-order heterogeneity parameters. In the study, less heterogeneous tumors, those with lower CoV, responded better to treatment. The heterogeneous distribution of the microsphere is due to altered hemodynamics of the hepatic artery blood flow and the heterogeneous vessel density inside the tumors [41]. Moreover, altered hemodynamics results from underlying tumor characteristics such as cellular proliferation, microvessel density, immune status, and hypoxia, which are relevant to tumor aggressiveness [20, 42]. Therefore, a high CoV indicates heterogeneous dose delivery to the tumor and may reflect the underlying tumor aggressiveness, resulting in a low therapeutic effect. While CoV is a relative metric, it showed a moderate inverse correlation with PET average dose (r = –0.367), suggesting that it reflects biologically relevant patterns in dose distribution rather than being dose-independent. From the study, the optimal cut-off of PET average dose to predict complete response was 471.2 Gy with a high specificity of 97%. Correspondingly, in a multicenter cohort to evaluate the dose-pathological correlation, complete pathological necrosis was found in all patients with tumor doses exceeding 400 Gy [43]. However, in the lower dose range, there remains a significant overlap in average tumor dose between responders and non-responders. In this range, V-V205, D70, and CoV showed added value over the PET average dose for predicting treatment response. Pre-Dose, estimated from the 99mTc-MAA SPECT/CT, was significantly lower compared with the average absorbed dose derived from post-therapeutic 90Y-PET/CT (P < 0.001). Pre-dose had the lowest ability to predict complete response among the dose metrics, and there was no significant association with L-PFS. This is consistent with previous reports and may be attributed to the fact that the preferential delivery of radioactive microspheres to hypervascular tumors cannot be adequately captured by pre-therapeutic imaging [44]. Furthermore, in terms of physical characteristics such as size and morphology, MAA differs from glass microspheres, so their biodistributions often do not match to each other. Alternatively, an attempt has been made to define the virtual tumor absorbed dose in a pre-therapeutic scan on the assumption that all administered microspheres are delivered to the tumor and assessed the predictive value [31]. Virtual tumor dose well predicted the CR in nodular HCC patients. Further research is needed to develop a dose metric that can predict treatment response in the pre-therapeutic scan. In this study, we have applied a multiple VSV kernel (Gy/MBq·s) convolution approach for voxel-based dosimetry. This approach enables faster personalized dosimetry than those using patient-specific Monte Carlo simulations and has shown accurate dosimetry results in nonuniform-density media [29]. Furthermore, by applying multiple VSVs that vary according to CT-derived tissue density, our method can account for intratumoral and peritumoral density heterogeneity. Unlike single VSV methods, which assume a uniform soft-tissue medium, the multiple VSV approach allows dose estimation to reflect spatial variations in physical density between tumor and normal liver tissue, thereby improving the physical accuracy of voxel-level dosimetry. In this study, all PET images were reconstructed using a consistent protocol (OSEM with 2 iterations and 21 subsets, incorporating TOF, PSF modeling, and a 5-mm Gaussian post-filter) to minimize reconstruction-related variability across patients. Although this setting corresponds to a relatively low equivalent iteration and moderate smoothing, it was selected to reduce image noise and maintain stability in voxel-level parameters such as the CoV. Prior phantom studies have demonstrated that increasing the number of iterations improves image contrast but also amplifies noise, leading to higher CoV, whereas stronger Gaussian smoothing suppresses noise but may blur small high-dose regions, thereby underestimating metrics like V-V205 [45–47]. When different FWHM was applied, CoV had small variation, indicating that it may be less affected by the image resolution [48]. The inclusion of TOF and PSF modeling in our reconstruction protocol is known to enhance quantitative accuracy by improving spatial resolution and count localization. Although CoV and V-V205 were not tested across different reconstruction settings in this study, future work should systematically evaluate their sensitivity to reconstruction parameters to support standardization efforts in 90Y PET dosimetry. Some limitations should be noted. Firstly, as a retrospective study, it has an inherent limitation. Still, responses were evaluated using strictly predefined criteria at regular intervals. Secondly, most of the patient included in this study were treated using the strategy of boosted radioembolization, as a result, the median PET average dose and median pre-Dose was 373.7 Gy and 231.0 Gy, respectively. However, the effectiveness of the high-dose radiation delivery has been shown in recent studies [5, 10]. In a prospective randomized trial, DOSISPHERE-01, the efficacy of a personalized dose intensification approach was compared with the standard method [10]. In the personalized dosimetry group, the mean tumor absorbed dose was 331.1 Gy, and the median OS survival doubled from 10.7 months to 26.6 months. Recently, the LEGACY study reported that 88% of the best response was achieved with a median perfused volume absorbed dose of 410 Gy [5]. Overall, TARE with a boosted dose seems to be an effective treatment strategy for HCC patients. Thirdly, the favorable clinical outcomes, including a high CR rate, should be interpreted with caution considering the patient population. In this study, 62.2% of patients were classified as BCLC stage A and 67.8% had single nodular HCC, indicating a relatively early and localized disease. Therefore, extrapolation of these findings to more advanced HCC stages, including BCLC B or C, may be limited. In addition, since the analysis was performed on a lesion-by-lesion basis using the largest tumor per patient, the results may not fully represent patient-level treatment outcomes, especially in those with multifocal disease. Fourth, for the precise tumor segmentation, careful co-registration of the baseline imaging with the posttreatment PET/CT is necessary, however, small misalignments were likely to present given the differences in technique. The alignment was verified for each tumor and manually adjusted to minimize the error. Furthermore, since we have included patients with tumors 3 cm or larger and analyzed the largest tumor, the impact of small misalignment would be minimal. Fifth, the follow-up period was highly variable among patients (range 3.2–53.4 months), resulting in a high proportion of censored data (79%). While this variability may introduce some bias and the findings should be interpreted with caution, there was no statistically significant difference in follow-up duration between patients with and without progression. Lastly, all patients were treated with glass microspheres, further studies in the resin microsphere cohort are needed to assure the generalizability of the results of our study. Conclusion The intratumoral dose distribution was assessed based on voxel-based dosimetry using the dose metrics of CoV, D70, and V-V205. Voxel-based dose metrics were significant predictors for treatment response in HCC patients undergoing TARE, even in the subgroup of tumors with a low average absorbed dose. This study elucidates the advantage of dose metrics, reflecting the intratumoral dose distribution, over average tumor absorbed dose for predicting local disease control in patients with HCC. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 1206 kb) Author contributions All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Minseok Suh. The first draft of the manuscript was written by Minseok Suh and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Funding This work was supported by SNUH Lee Kun-hee Child Cancer & Rare Disease Project, Republic of Korea (24A-023–0100), Technology development Program (RS-2025–02222237) funded by the Ministry of SMEs and Startups, National Research Foundation of Korea grants funded by the Korean government (MSIT) (RS-2022-00164776), and Seoul R&BD Program (BT240153) through the Seoul Business Agency (SBA) funded by The Seoul Metropolitan Government. Data availability The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval All procedures performed in studies were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. All included data were collected as study protocol approved by the local institutional ethics committee (approval No. H-2207–030-1337). Consent to participate This retrospective study was approved by the institutional review board in accordance with the ethical standard and with the Helsinki Declaration of 1964 and its later amendments. The requirement for written informed consent was waived. Consent for publication N/A. Competing interests No other potential conflicts of interest relevant to this article exist. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Jin Chul Paeng, Email: paengjc@snu.ac.kr. Hyo-Cheol Kim, Email: radioembolization@snu.ac.kr. References Associated Data Supplementary Materials Supplementary file1 (DOCX 1206 kb) Data Availability Statement The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.