Fully Automated CT-Based Quantitative Body Composition Analysis for Predicting Survival in Patients with HCC Undergoing TACE: A Dual-Cohort Study

Fully Automated CT-Based Quantitative Body Composition Analysis for Predicting Survival in Patients with HCC Undergoing TACE: A Dual-Cohort Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Fully Automated CT-Based Quantitative Body Composition Analysis for Predicting Survival in Patients with HCC Undergoing TACE: A Dual-Cohort Study Krzysztof Bartnik, Belgutei Tsevegmed, Kamil Książek, Zuzanna Wojtczak, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8713867/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 10 You are reading this latest preprint version Abstract Transarterial chemoembolization (TACE) is a standard treatment for patients with unresectable hepatocellular carcinoma (HCC), yet existing models provide limited individualized risk stratification. Automated CT-derived body composition analysis has emerged as an objective marker of patient physiological reserve, but its value in prognostication in TACE patients is insufficiently studied. Therefore, the aim of the study was to evaluate the prognostic value of a fully automated, open-source pipeline for CT-based body composition analysis in predicting overall survival (OS) in patients with HCC undergoing TACE. In this study, we used two independent cohorts of treatment-naive patients undergoing TACE: the WAW-TACE cohort (development) (n = 230) and the HCC-TACE-Seg cohort (validation) (n = 100). Skeletal muscle and fat metrics were extracted from pre-treatment CTs using a standardized deep learning pipeline and normalized by sex. Survival analyses were performed using Cox proportional hazards (CoxPH) models and random survival forests (RSF). Skeletal muscle density (SMD) at the L3 level was the strongest and independent predictor of OS across both cohorts (HR: development, 0.84; p = 0.029; validation, 0.79; p = 0.028). This association remained significant after adjustment for the best-performing clinical composite scores: mHAP-2 in the development (adjusted HR = 0.68; p = 0.049) and CLIP in the validation cohort (adjusted HR = 0.43; p = 0.003). In CoxPH, the addition of SMD metrics resulted in only modest improvements in discrimination (ΔC-index 0.011–0.037) that did not reach statistical significance. In contrast, RSF analysis demonstrated a statistically significant improvement in model discrimination when muscle-based variables were added to clinical features (ΔC-index = 0.023; p < 0.001). Fully automated CT-derived body composition metrics were independently associated with OS in patients treated with TACE. While their addition to traditional clinical models resulted in only modest improvements in discrimination, muscle density provided complementary prognostic information. Notably, statistically significant increases in model performance were observed when muscle-based features were integrated into RSF, highlighting their potential value within machine learning frameworks. Table of Contents Health sciences/Biomarkers Biological sciences/Cancer Health sciences/Medical research Health sciences/Oncology Health sciences/Risk factors Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Transarterial chemoembolization (TACE) is a first line, minimally invasive procedure performed primarily in patients with unresectable hepatocellular carcinoma (HCC) [ 1 , 2 ]. Most prognostic scores created for HCC mostly rely on laboratory evaluations of underlying liver disease combined with basic radiologic characteristics, such as the number of tumor lesions and their size [ 3 , 4 ]. Nevertheless, these models are unable to adequately represent the diversity of HCC patients and have limited predictive accuracy [ 3 , 5 ]. Even while two patients may present with tumors of similar size and morphology as well as comparable laboratory parameters, their entire physiological state and, thus, prognosis may differ significantly [ 6 ]. As a result, there is a need for more individualized pre-treatment risk-stratification tools that can optimize treatment strategies. In recent years, the significance of individualized prognostication has been emphasized by a growing number of studies [ 7 , 8 ]. This concept is reflected in the growing interest in frailty-based prognostication approaches [ 9 ]. Frailty is understood as reduced physiologic reserve and greater vulnerability to unfavorable treatment outcomes [ 10 ]. In clinical practice, frailty is commonly assessed using subjective measures, such as the Clinical Frailty Scale (CFS), which, despite its widespread use, is limited by inter-observer variability [ 11 , 12 ]. In turn, objective body composition analysis has recently been suggested as a scalable alternative for conventional frailty criteria like age or performance status [ 13 ]. Variables such as skeletal muscle volume, visceral fat area, and other morphometric indices have been proposed as significant indicators of outcomes in a variety of clinical contexts, including overall survival (OS) in patients with HCC or colorectal cancer [ 14 ]. Conventionally, the quantification of these factors relies on the laborious process of manually segmenting regions of interest, which requires hand-crafted input and potentially introduces bias [ 15 ]. Therefore, automated segmentation models have gained increasing attention. However, most of the studied models are not open access and are tailored to specific populations, which importantly restricts their wider clinical usage [ 16 ]. A recently published model from Chen et al. enables standardized and fully automated segmentation and extraction of imaging features of muscle and adipose tissue volumes on computed tomography (CT) images, providing a comprehensive body composition profile [ 13 , 17 ]. The pipeline was developed, standardized, and validated on a heterogeneous cohort of CT scans representing multiple patient populations, ensuring generalizability. Recent studies on colorectal cancer patients have demonstrated that features extracted using this method can be utilized for survival prediction and treatment-outcome assessment [ 14 ]. However, the prognostic value of automatically derived body composition parameters in patients with HCC, particularly in the setting of TACE, remains insufficiently explored. Therefore, the aim of our study was to evaluate whether body composition parameters derived using a standardized and automated approach may serve as useful prognostic markers for prediction of survival in patients with HCC treated with TACE. For this purpose, we applied a recently released an open-access access model for automated body composition analysis on CT imaging in two independent HCC cohorts. 2. Materials and Methods The study was conducted in accordance with the Declaration of Helsinki. Institutional review board approval and informed consent were waived, as the study used fully anonymized, publicly available open-science datasets. 2.1 Patient cohorts In this retrospective study, we utilized two independent cohorts: WAW-TACE and HCC-TACE-Seg. The WAW-TACE dataset includes 233 patients treated between 2016 and 2021 [ 18 ]. All patients were treatment-naive and diagnosed with unresectable HCC, subsequently managed with TACE monotherapy. Baseline imaging consists of multiphase contrast-enhanced CT acquired within 90 days prior to the initial TACE cycle. In turn, the HCC-TACE-Seg dataset features 105 patients treated between 2002 and 2012 [ 19 ]. All cases from this dataset represent confirmed HCC managed with TACE as the sole first-line therapy or as initial bridging therapy. Baseline imaging includes multiphase contrast-enhanced CT performed 1–12 weeks (average 3 weeks) before the first TACE session. 2.2 Clinical variables The clinical variables analyzed included the following: 1) For WAW-TACE dataset: age, sex, alpha-fetoprotein (AFP), total bilirubin, serum albumin, number of tumor lesions, tumor size, Child-Pugh score (CPS), serum creatinine, alanine aminotransferase (ALT), Barcelona Clinic Liver Cancer stage (BCLC), and international normalized ratio (INR). In addition, the following clinical scores were evaluated: ALBI score, mHAP II score, 6&12 score, and HAP score [ 3 , 20 – 22 ]. 2) For HCC-TACE-Seg dataset: sex, age, CPS, AFP, BCLC, and numeric clinical staging scores including CLIP (Cancer of the Liver Italian Program) score, Okuda, TNM (Tumor-Node-Metastasis). 2.3 Extraction of image-based body composition metrics All segmentations and body composition feature extractions were performed using the portal venous phase of the pre-TACE CT examinations. For the extraction of body composition metrics, we utilized the open-source pretrained deep learning pipeline by Chen et al [ 13 ]. The authors designed an automated, deep learning-based framework capable of segmenting skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue on axial CT scans of the abdomen and pelvis. The pipeline supports both 2D analysis at the L3 level and 3D volumetric assessment from T12 to L4, providing a comprehensive and standardized set of quantitative measurements, as described in the study by Hanxue et al [ 14 ]. The extracted body composition metrics included the following variables. For 2D measurements at the L3 level, these comprised skeletal muscle area (mm²), skeletal muscle density (2D SMD) [HU], subcutaneous fat area (mm²), visceral fat area (mm²), intramuscular fat area (mm²), total fat area (mm²), and total body cross-sectional area (mm²). For 3D measurements between T12 and L4, the extracted variables included skeletal muscle volume (mm³), skeletal muscle density (3D SMD) [HU], subcutaneous fat volume (mm³), visceral fat volume (mm³), intramuscular fat volume (mm³), total fat volume (mm³), total body volume (mm³), and craniocaudal body height between T12 and L4 (mm). 2.4 Outcome variables The primary endpoint of the study was to evaluate the association between automated body composition variables and OS in HCC patients undergoing TACE. Secondary endpoints included 3-year mortality rates and the assessment of whether the addition of CT-derived body-profiling variables could significantly improve the predictive performance of state-of-the-art clinical scores, as well as the best-performing combination of clinical variables. 2.5 Statistical analysis Quantitative body composition metrics were standardized using sex-specific Z-score normalization to control for sex-related physiological differences that were suggested by previous studies [ 25 ]. For each variable, sex-specific means and standard deviations were calculated, and normalized values were derived as Z-scores according to the formula: \(\:Z=\frac{x-{\mu\:}_{sex}}{{\sigma\:}_{sex}}\) where \(\:x\:\) denotes the observed value, while \(\:{\mu\:}_{sex}\) and \(\:{\sigma\:}_{sex}\) represent the sex-specific mean and standard deviation, respectively. 2D SMD was dichotomized using a data-driven cut-off, resulting in a binary muscle density indicator (binary 2D SMD). The classification into low and high subgroups in the Kaplan-Meier analysis was performed using an optimal cut-off value, determined by maximizing the log-rank test statistic (Contal & O’Quigley-like method) [ 26 ]. Body-composition parameters and clinical variables were first evaluated using univariable Cox proportional hazards (CoxPH) analyses. Variables that reached statistical significance (p < 0.05) in the univariable analysis were subsequently included in multivariable CoxPH models to assess their independent predictive value. The concordance index (C-index) was used to evaluate the predictive performance of the CoxPH models for OS. Additionally, the area under the ROC curve (AUC) was calculated for 3-year mortality. Performance differences between models were quantified using ΔC-index and ΔAUC(t), with bootstrap-based confidence intervals. Additionally, a machine learning-based approach was used to assess the impact of body composition parameters on OS prediction. Firstly, a 5-fold cross-validation (CV) experiment using Random Survival Forests (RSF) with all available variables and 100 estimators was performed. In each cross-validation step, a permutation-based feature importance was calculated. After five CV folds, variables having a positive mean importance score, i.e. causing a more accurate C-index estimation, were selected. Then, the RSF model with variables considered important in the previous step (both clinical and body composition) was compared with the model trained using only a significant subset of clinical variables. This evaluation was performed in another 5-fold cross-validation experimental scenario with 10 repetitions. Finally, to assess the significance of the difference in model performance, a t-test for related samples was performed. All statistical analyses and figure generation were performed in Python (Python Software Foundation 2024, Language Reference Version 3.13.0). Data handling was conducted with the pandas and NumPy libraries. Survival analyses were performed using the lifelines and scikit-survival packages, including Cox proportional hazards regression ( CoxPHFitter ) and calculation of Harrell’s concordance index. Standard Python utilities ( os , pathlib ) and typing modules were used for file handling and code organization. Analyses were executed in a reproducible Python environment, and the complete source code and environment specifications can be provided upon reasonable request. 3. Results 3.1 Characteristics of patient populations WAW-TACE dataset: the final analysis included 230 patients who underwent TACE and had an available portal venous phase. OS was 28.6 months (95% CI: 21.8–33.8); 1-year and 3-year mortality rates were 26.5% and 59.7%, respectively. HCC-TACE-Seg dataset: 100 patients who underwent TACE and had an available portal venous phase were included. OS was 24.0 months (95% CI: 19.6–29.0). The 1-year and 3-year mortality rates were 22.1% and 73.6%, respectively. A detailed characterization of both evaluated cohorts is presented in Table 1 . Table 1 Baseline clinical characteristics of patients included in the study in the WAW-TACE and HCC-TACE-Seg cohorts. Continuous variables are reported as medians and categorical variables as counts with percentages. Differences between cohorts were assessed using appropriate statistical tests for continuous and categorical variables, with corresponding p values reported. Variable WAW-TACE (n = 230) HCC-TACE-Seg (n = 100) p value Female sex, n (%) 48 (20.9) 35 (35.0) 0.01 Age, median (years) 66 68.5 0.17 Child-Pugh class, n (%) 0.08 A 204 (88.7) 81 (81.0) B 26 (11.3) 18 (18.0) C - 1 (1.0) AFP, n (%) 0.46 < 400 ng/mL 176 (76.5) 72 (72.0) ≥ 400 ng/mL 54 (23.5) 28 (28.0) AFP, median (ng/mL) 14.25 32.9 0.15 BCLC stage, n (%) < 0.01 0/A 168 (73.0) 15 (15.0) B 62 (27.0) 23 (23.0) C - 62 (62.0) 3.2 WAW-TACE 3.2.1 Univariable survival analysis Full univariate CoxPH results for all raw and sex-normalized fat and muscle-based predictors are provided in the Supplementary Tables S1 and S2, respectively. All subsequent muscle- and fat-related variables reported in the following sections refer to these sex-normalized measures. Among all image-derived variables, only 2D SMD and 3D SMD were significant predictors of OS. The corresponding hazard ratios were 0.84 (95% CI, 0.72–0.98; p = 0.029) for 2D SMD and 0.85 (95% CI, 0.73-1.00; p = 0.045) for 3D SMD. Among clinical variables, serum albumin, tumor size, log-transformed AFP, total bilirubin, number of tumor lesions, and Child-Pugh score were significant predictors of OS (Supplementary Table S3). Among all variables evaluated in the univariable analysis, the mHAP-2 score was the strongest predictor of OS (HR 1.58, 95% CI 1.32–1.89, p < 0.01) with the lowest C-index of 0.641 (Supplementary Table S3). 3.2.2 Determination of binary indicator for skeletal muscle density at the L3 level In the cut-off selection procedure for 2D SMD, the optimal threshold was identified as -0.85 with the log-rank test p < 0.001 (Fig. 3 ). This binary SMD indicator was subsequently evaluated in univariate survival models. For OS, the binary SMD was significantly associated with mortality risk (HR = 0.53, 95% CI: 0.37–0.76, p = 0.0005). For 3-year mortality, the association was also significant (HR = 0.59, 95% CI: 0.40–0.86, p = 0.006). Patients below the cut-off had a median OS of 19.9 months (95% CI: 9.3–32.8), whereas those above the cut-off had a median OS of 32.3 months (95% CI: 22.9–38.4). At 3 years, survival rates were 22.8% (95% CI: 11.8–36.0) and 44.5% (95% CI: 37.2–51.5), respectively. 3.2.3 Multivariable analysis After adjusting for the mHAP-2 score (the strongest clinical composite score associated with OS), binary 2D SMD remained independently associated with OS (HR 0.68, 95% CI 0.46-1.00, p = 0.049). When evaluated as a predictive marker, adding binary 2D SMD to the mHAP-2 model resulted in only a small improvement in discrimination: the C-index increased from 0.641 to 0.657 (Δ = 0.015; 95% CI -0.006 to 0.039; p = 0.513). A similar modest gain was observed when adding continuous 2D SMD instead of the binary cut-off (C-index 0.657; Δ = 0.016; 95% CI -0.001 to 0.033; p = 0.485). Similar results were obtained in an analysis adjusting for individual clinical variables rather than composite clinical scores. Binary 2D SMD remained an independent predictor after adjustment for state-of-the-art clinical confounders identified in univariable analysis (Supplementary Table S4), including serum albumin, total bilirubin, log-transformed AFP, Child-Pugh score, tumor size, and number of tumor lesions. Incorporation of binary 2D SMD to clinical variables resulted in a modest improvement in model discrimination, with the concordance index increasing from 0.672 for the clinical model to 0.684 for the extended model (ΔC-index = 0.011, 95% CI − 0.003 to 0.027; bootstrap p = 0.50). A similar pattern was observed in time-specific discrimination. At 3 years, the time-dependent AUC increased from 0.698 for the clinical-only model to 0.720 after inclusion of binary 2D SMD (ΔAUC = 0.022); however, this improvement did not reach statistical significance (95% CI, − 0.002 to 0.046; bootstrap p = 0.50). A comparable improvement was observed when a continuous 2D SMD was used instead (ΔAUC = 0.026, p = 0.51). 3.2.4 Machine learning-based feature importance estimation The feature importance estimation with RSF revealed that 9 clinical and 3 muscle-based variables are valuable predictors with positive importance scores (Fig. 4 ). The most important variables were log-transformed AFP, total bilirubin, and tumor size, belonging to the category of clinical variables. The three selected muscle-based predictors, i.e. skeletal muscle area at the L3 level, 2D SMD, and skeletal muscle volume, were considered to have a positive influence on model estimations. The remaining 15 variables were excluded from further calculations. In the second cross-validation stage, a model with a subset of all variables was compared with a model having only a subset of clinical predictors. The combined model obtained the C-index of 0.6578, while the clinical-only model achieved 0.6352. Since the differences in predictions between the two models were normally distributed according to the results of the Shapiro-Wilk test (W = 0.98, p = 0.59), the one-sided t-test for related samples was calculated. Its results (t = 6.04, p < 0.001) indicate that the inclusion of muscle-based predictors together with clinical variables resulted in a statistically significant improvement in model performance compared with the clinical-only model. 3.3 HCC-TACE-Seg 3.3.1 Univariable survival analysis Full univariate CoxPH results for all sex-normalized fat and muscle metrics are provided in the Supplementary Table S5. Among all evaluated factors, 2D SMD and 3D SMD were the only variables significantly associated with OS in the univariable analysis, which was analogous to results on the WAW-TACE dataset. The corresponding hazard ratios were 0.79 (95% CI, 0.65–0.98; p = 0.028) for 2D SMD and 0.80 (95% CI, 0.65–0.99; p = 0.045) for 3D SMD. 3.3.2 Determination of binary indicator for skeletal muscle density at the L3 level In the cut-off selection procedure for 2D SMD, the optimal threshold was identified as 0.61, which yielded a statistically significant separation in survival curves (log-rank p = 0.002). The dichotomized variable was subsequently evaluated in univariable CoxPH, demonstrating a strong association with overall survival (HR = 0.43, 95% CI: 0.25–0.75, p = 0.0029). The median OS in the cohort was 24.0 months (95% CI: 19.6–29.0). Patients classified as having low 2D SMD according to the binary indicator had a median OS of 24.0 months (95% CI: 17.5–28.1), whereas those classified as having high 2D SMD had a median OS of 35.6 months (95% CI: 15.2–81.6). At 3 years, OS was 18.6% (95% CI: 10.6–28.5) and 48.0% (95% CI: 27.8–65.6), respectively. 3.3.3 Multivariable analysis Within this dataset, the CLIP score emerged as the strongest clinical predictor of mortality (HR 1.43, 95% CI 1.15–1.77, p = 0.0012) (Supplementary Table S6). In the multivariable CoxPH model, including the CLIP score and binary 2D SMD, both covariates were significantly associated with OS. Binary 2D SMD showed a hazard ratio of 0.42 (95% CI: 0.24–0.73, p = 0.002). In the multivariable CoxPH analysis for OS, incorporation of the binary 2D SMD into the clinical model based on the CLIP score resulted in a numerically higher discriminative performance. Specifically, the C-index increased from 0.565 for the clinical-only model to 0.602 for the extended model including the binary 2D SMN, corresponding to an absolute ΔC-index of 0.037. However, this improvement did not reach statistical significance (95% CI, − 0.018 to 0.094; bootstrap p = 0.498). At 3 years, time-dependent discrimination improved from an AUC of 0.579 to 0.704 (ΔAUC = 0.125). While the bootstrap confidence interval suggested a non-zero effect (95% CI, 0.019–0.234), bootstrap-based hypothesis testing did not indicate statistical significance (p = 0.466). 4. Discussion In this study, we evaluated the prognostic significance of standardized CT-derived body composition parameters on survival in patients with unresectable HCC treated with TACE. The results were validated in two independent datasets, supporting their robustness and generalizability, which is a major strength of this study. Despite differences in sample size, timeframe, and patient characteristics across both cohorts, skeletal muscle-based metrics consistently were significant and independent predictors of OS. The reproducibility of this finding across two different clinical populations suggests that muscle quality captures important aspects of physiological reserve and patient frailty that traditional clinical scores do not fully capture. Specifically, in the WAW-TACE cohort, 2D SMD and 3D SMD demonstrated an association with survival, and the effect of 2D SMD remained preserved even after adjustment for established clinical scores such as the mHAP-2 score. Similarly, in the HCC-TACE-Seg dataset, both muscle density measurements (2D SMD and 3D SMD) were significantly associated with mortality risk, reinforcing the robustness of this metric across different clinical settings. At the same time, our evaluation of predictive performance showed that although muscle density remained independently associated with mortality in multivariable CoxPH models, its incremental contribution to overall model discrimination was modest. For example, in the WAW-TACE cohort, adding binary 2D SMD to a combination of significant clinical variables increased the C-index only slightly (ΔC-index = 0.011), and a similar limited improvement was observed in the 3-year AUC (ΔAUC = 0.022). Although the inclusion of muscle density consistently yielded numerically higher C-indices, the bootstrap CI crossed zero and the differences were not statistically significant, indicating that the observed improvements in discrimination were modest and likely reflect a small incremental prognostic value. These findings further suggest that the study may be underpowered to detect small but statistically significant differences in model performance. Importantly, subsequent machine learning evaluation further supported these findings, as demonstrated by the RSF analysis, in which a model incorporating a subset of both clinical and muscle-based variables achieved a significantly higher performance than the clinical-only model (ΔC-index = 0.023; p < 0.001). These findings indicate that automated CT-based body composition analysis including muscle density assessment provides prognostically relevant information not fully captured by standard clinical variables, but the magnitude of improvement in model performance remains relatively small. In other words, automated skeletal muscle density may be valuable as a complementary imaging-derived biomarker within broader multimodal risk-stratification frameworks, rather than as a standalone predictor intended to substantially enhance discrimination. In particular, muscle density may be especially useful in machine learning-based and radiomics-driven approaches. Those models frequently rely on multiple imaging features, where SMD can serve as an additional biologically interpretable variable contributing incremental prognostic value and improving overall model performance [ 27 ]. Its integration into imaging-only models warrants further investigation. Our results align with previous reports linking reduced muscle quality to poorer outcomes following colorectal surgery. A recent study by Hanxue Gu et al. showed that sex-normalized SMD -measured in the same way as in our study - was a strong and independent predictor of 1-year mortality after colectomy (OR 0.42) [ 14 ]. The authors showed that these metrics have prognostic value independent of traditional clinical variables and enhance the discriminatory accuracy of established risk calculators (ΔC-index = 0.07, 0.80 vs. 0.73), whereas in our study, they improved prediction only modestly (ΔC-index values ranging from 0.011 to 0.037). Importantly, our similar observations in an entirely different clinical context support the idea that automated SMD represents a reliable biomarker that generalizes across a variety of clinical scenarios and imaging protocols. This implies that automated CT-based muscle quality assessment captures significant parts of patient frailty and physiologic reserve that are not represented by conventional clinical variables. As a result, these measures could be useful as adjuncts for individualized risk assessment in a variety of patient groups. This method is easy to implement and does not require additional procedures beyond standard clinical care. Once included in the clinical workflow, the model can be fully automated. These findings are consistent with a recent meta-analysis that summarizes the predictive significance of body composition in HCC patients. The pooled HR = 1.38 for sarcopenia as a predictor of OS in TACE patients, suggesting that subjects with low muscle quality have a significantly higher mortality [ 15 ]. However, the methodologies used to assess muscle mass and muscle quality differed substantially across the included studies, ranging from manual and semi-automated segmentations at different anatomical levels to threshold-based approaches and ROI measurements. This variability makes direct comparison more difficult, restricts wider clinical translation, and underscores the need for standardized, reproducible approaches to body composition evaluation. In this context, the fully automated and publicly available model used in our study offers a potential solution by providing consistent and observer-independent measurements that may facilitate standardization across centers and imaging protocols. A conceptually similar automated approach was previously presented by Wang et al., who developed an in-house deep convolutional neural network to facilitate psoas muscle segmentation on CT studies [ 16 ]. Their study served primarily as a proof-of-concept that deep learning can replace manual measurements of psoas muscle volume and that such an automated psoas muscle index is independently associated with OS in patients with liver cirrhosis. A cohort of 254 people with chronic liver disease was used for clinical validation in their investigation. The study showed that automated psoas size measurement was a significant survival predictor even after adjusting for age, sex, and liver function. Specifically, the CT-derived psoas muscle index was significantly associated with survival (HR = 0.86), but the study did not present parametric model performance metrics comparing models with and without this parameter. Moreover, the model used in that study was trained and assessed mostly on internal datasets, lacked external validation, and was not made available as a standardized, publicly accessible tool, that restricts its wider clinical applicability. Despite limitations, these results offered proof that prognostic information is carried by image-based muscle measures. Our study expands this concept beyond chronic liver disease in general to a more clinically specific context: TACE patients with HCC. Additionally, our analysis relied on an open-source, pretrained, fully automated pipeline that, once applied, does not require manual annotation or additional radiologist input, unlike Wang et al.'s approach. This substantially lowers the barrier to integration in other clinical centers and supports the feasibility of using standardized body-composition profiling as part of routine pre-TACE risk stratification. A series of related observations was reported by Surov et al., who studied the prognostic role of muscle quality in patients with HCC treated with different locoregional therapies. In one study, involving patients undergoing combined SIRT and sorafenib treatment, skeletal muscle density was measured semi-automatically at the L3 level with manual radiologist corrections [ 28 ]. Myosteatosis was associated with OS in the univariable analysis, although this effect did not remain significant after adjustment for demographic and clinical variables. Notably, subgroup analyses showed that myosteatosis was a significant predictor only in patients with alcohol-related cirrhosis, suggesting that its importance may depend on underlying liver disease. In another study by the same authors, using a similar semi-automated methodology in cirrhotic patients treated with TACE, myosteatosis again did not demonstrate an independent association with survival, whereas the albumin-muscle density score emerged as a significant predictor [ 29 ]. Both studies reported associations with survival but did not include a formal evaluation of the additive model performance of imaging-derived muscle parameters beyond hazard ratios. The broader prognostic relevance of body composition measurements has been demonstrated across multiple clinical settings. Sarcopenia and myosteatosis have been associated with increased mortality in patients with breast cancer, chronic liver disease, chronic kidney disease, and liver transplantation [ 30 – 34 ]. Collectively, these results suggest that muscle mass and muscle quality are significant indicators of physiological susceptibility in oncology and other medical scenarios. This supports the idea that a fully automated, standardized CT-based methodology, like the one employed in our study, could be potentially applicable in a variety of illness scenarios that require individualized risk stratification. 4.1 Limitations This study has several limitations that should be acknowledged. First, the retrospective design of both cohorts introduces the possibility of selection bias and unmeasured confounding that may influence outcomes independently of body composition metrics. Second, although the use of two independent datasets enhances the generalizability of our findings, the cohorts differed with respect to treatment timeframe (2002–2012 vs. 2016–2021), clinical practice standards, imaging protocols, scanner types, and baseline patient characteristics. These sources of heterogeneity, combined with the lack of scanner-specific normalization, may affect the comparability of absolute muscle-density values and the magnitude of observed associations. Finally, the optimal cut-off for muscle density was derived using maximally selected log-rank statistics, which carries an inherent risk of overfitting. Data-driven thresholds may overestimate effect sizes unless validated prospectively or in an independent test set. 5. Conclusions In patients with unresectable HCC treated with TACE, automated, quantitative CT-based assessment of skeletal muscle density emerged as an independent predictor of OS. This effect was reproducible in two independent cohorts and consistent with findings from an external colorectal cancer study, suggesting broader relevance beyond HCC. Although its addition led to only modest improvements in discrimination over established clinical models, the automated muscle-density measure is easy to implement, scalable, and may serve as a useful complementary quantitative imaging biomarker within multimodal risk-stratification frameworks. Declarations Funding This project was supported by the SONATINA grant from the National Science Centre (NCN), No. 2025/56/C/NZ7/00276 (DEC-2025/56/C/NZ7/00276). Author contributions statement Study design: KB; KK Methodology: KB, KK Statistical analysis: KB, KK, ZW Segmentation and feature extraction: BT Visualization: KB, KK, ZW Supervision: PB Writing–Original Draft Preparation: KB Writing–Review, Feedback & Editing: all authors Approval of final version of submitted manuscript: all authors. Additional information The authors declare no competing interests. Author Contribution Study design: KB, KK; Methodology: KB, KK; Statistical analysis: KB, KK, ZW; Segmentation and feature extraction: BT; Visualization: KB, KK, ZW; Supervision: PB; Writing–Original Draft Preparation: KB; Writing–Review, Feedback & Editing: all authors; Approval of final version of submitted manuscript: all authors. Data Availability All datasets used in this study are publicly available. The WAW-TACE dataset is available in the Radiology: Artificial Intelligence under DOI: 10.1148/RYAI.240296. The HCC‑TACE‑Seg dataset is available in the Scientific Data under DOI: 10.1038/s41597‑023‑01928‑3. References Marrero, J. A. et al. Diagnosis, Staging, and Management of Hepatocellular Carcinoma: 2018 Practice Guidance by the American Association for the Study of Liver Diseases. Hepatol. Hepatol. 68 (2), 723–750. 10.1002/HEP.29913 (2018). Heimbach, J. K. et al. AASLD guidelines for the treatment of hepatocellular carcinoma. Hepatology Vol. 67, 358–380 (John Wiley and Sons Inc., 2018). 1. Wang, Q. et al. Development of a prognostic score for recommended TACE candidates with hepatocellular carcinoma: A multicentre observational study. J. Hepatol. J. Hepatol. 70 (5), 893–903. 10.1016/J.JHEP.2019.01.013 (2019). Johnson, P. J. et al. Assessment of liver function in patients with hepatocellular carcinoma: A new evidence-based approach - The albi grade. J. Clin. Oncol. Am. Soc. Clin. Oncol. 33 (6), 550–558. 10.1200/JCO.2014.57.9151 (2015). Pinato, D. J. et al. Combined sequential use of HAP and ART scores to predict survival outcome and treatment failure following chemoembolization in hepatocellular carcinoma: A multi-center comparative study. Oncotarget Impact Journals LLC . 7 (28), 44705–44718. 10.18632/oncotarget.9604 (2016). Kelley, R. K. & Greten, T. F. Hepatocellular Carcinoma - Origins and Outcomes. New England Journal of Medicine. Mass. Med. Soc. 385 (3), 280–282 (2021). doi: 10.1056/NEJMCIBR2106594;ISSUE:ISSUE:DOI. Vitale, A., Trevisani, F., Farinati, F. & Cillo, U. Treatment of Hepatocellular Carcinoma in the Precision Medicine Era: From Treatment Stage Migration to Therapeutic Hierarchy. Hepatology: Official Journal of the American Association for the Study of Liver Diseases. John Wiley and Sons Inc; ;72(6):2206–2218. doi: 10.1002/HEP.31187 . (2020). Vitale, A. et al. Personalised management of patients with hepatocellular carcinoma: a multiparametric therapeutic hierarchy concept. Lancet Oncol. Elsevier Ltd. 24 (7), e312–e322. 10.1016/S1470-2045(23)00186-9 (2023). Hamaker, M. E. et al. Frailty screening methods for predicting outcome of a comprehensive geriatric assessment in elderly patients with cancer: A systematic review. Lancet Oncol. Elsevier . 13 (10), e437–e444. 10.1016/S1470-2045(12)70259-0 (2012). Ethun, C. G. et al. Frailty and cancer: Implications for oncology surgery, medical oncology, and radiation oncology. CA Cancer J. Clin. Wiley . 67 (5), 362–377 (2017). .21406;REQUESTEDJOURNAL:JOURNAL:15424863;WGROUP:STRING:PUBLICATION. Welford, J. et al. The Clinical Frailty Scale can indicate prognosis and care requirements on discharge in oncology and haemato-oncology inpatients: A cohort study. Eur. J. Cancer Care (Engl) John Wiley Sons Inc . 31 (6), e13752. 10.1111/ECC.13752 (2022). Shah, D. et al. Use of the Rockwood Clinical Frailty Scale in patients with advanced hepatopancreaticobiliary malignancies. Expert Rev. Anticancer Ther. Expert Rev. Anticancer Ther. 22 (9), 1009–1015. 10.1080/14737140.2022.2096594 (2022). Chen, Y. et al. Automated Muscle and Fat Segmentation in Computed Tomography for Comprehensive Body Composition Analysis. Machine Learning for Biomedical Imaging. Mach. Learn. Biomedical Imaging . 3 (November 2025), 581–617. 10.59275/J.MELBA.2025-AG5G (2025). Gu, H. et al. Improving Surgical Risk Prediction Through Integrating Automated Body Composition Analysis: a Retrospective Trial on Colectomy Surgery. ; (2025). https://arxiv.org/pdf/2506.11996 . Accessed December 11, 2025. Wang, A. et al. Impact of body composition on the prognosis of hepatocellular carcinoma patients treated with transarterial chemoembolization: A systematic review and meta-analysis. Heliyon Vol. 10, e25237 (Elsevier Ltd, 2024). 310.1016/J.HELIYON.2024.E25237 Wang, N. C. et al. Automated Measurements of Muscle Mass Using Deep Learning can Predict Clinical Outcomes in Patients with Liver Disease. Am. J. Gastroenterol. Wolters Kluwer Health . 115 (8), 1210. 10.14309/AJG.0000000000000662 (2020). GitHub - mazurowski-lab/CT-Muscle-and-Fat-Segmentation. (2025). https://github.com/mazurowski-lab/CT-Muscle-and-Fat-Segmentation . Accessed December 11. Bartnik, K. et al. WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data. Radiol. Artif. Intell. Radiological Soc. North. Am. Inc . 6 (6). 10.1148/RYAI.240296/ASSET (2024). /IMAGES/LARGE/RYAI.240296.FIG3.JPEG Moawad, A. W. et al. Multimodality annotated hepatocellular carcinoma data set including pre- and post-TACE with imaging segmentation. Sci. Data Nat. Publishing Group. ; ; 10 (1). doi: 10.1038/S41597-023-01928-3 . (2023). Kadalayil, L. et al. A simple prognostic scoring system for patients receiving transarterial embolisation for hepatocellular cancer. Annals of Oncology. Ann. Oncol. 24 (10), 2565–2570. 10.1093/annonc/mdt247 (2013). Park, Y. et al. Addition of tumor multiplicity improves the prognostic performance of the hepatoma arterial-embolization prognostic score. Liver Int. Liver Int. 36 (1), 100–107. 10.1111/LIV.12878 (2016). Lee, I. C. et al. A new ALBI-based model to predict survival after transarterial chemoembolization for BCLC stage B hepatocellular carcinoma. Liver Int. Liver Int. 39 (9), 1704–1712. 10.1111/LIV.14194 (2019). Daniele, B. et al. Cancer of the Liver Italian Program (CLIP) score for staging hepatocellular carcinoma Vol. 37, S206–S209 (Hepatology Research. John Wiley & Sons, Ltd, 2007). s210.1111/J.1872-034X.2007.00186.X Kinoshita, A. et al. Staging systems for hepatocellular carcinoma: Current status and future perspectives. World J. Hepatol. Baishideng Publishing Group. Co. 7 (3), 406. 10.4254/WJH.V7.I3.406 (2015). Meister, F. A. et al. The Role of Sarcopenia and Myosteatosis in Short- and Long-Term Outcomes Following Curative-Intent Surgery for Hepatocellular Carcinoma in a European Cohort. Cancers (Basel) Cancers (Basel) . 14 (3). 10.3390/CANCERS14030720 (2022). Contal, C. & O’Quigley, J. An application of changepoint methods in studying the effect of age on survival in breast cancer. Comput. Stat. Data Anal. N.-Holl. 30 (3), 253–270. 10.1016/S0167-9473(98)00096-6 (1999). Bartnik, K. et al. A novel radiomics approach for predicting TACE outcomes in hepatocellular carcinoma patients using deep learning for multi-organ segmentation. Scientific Reports Nature Publishing Group; 2024;14(1):1–13. (2024). 14:1 10.1038/s41598-024-65630-z Surov, A. et al. Skeletal muscle quality predicts overall survival in advanced liver hepatocellular carcinoma treated with SIRT and sorafenib: A subanalysis of the SORAMIC trial. United Eur. Gastroenterol. J. United Eur. Gastroenterol. J. 12 (8), 1016–1027. 10.1002/UEG2.12627 (2024). Surov, A. et al. Albumin-muscle density score predicts overall survival in patients with hepatocellular cancer undergoing treatment with transarterial chemoembolization. J. Cancer Res. Clin. Oncol. J. Cancer Res. Clin. Oncol. 150 (12). 10.1007/S00432-024-06043-3 (2024). Gourd, E. Sarcopenia and adiposity linked to overall survival. Lancet Oncol. NLM (Medline) . 19 (5), e239. 10.1016/S1470-2045(18)30284-5 (2018). Liu, C. et al. Sarcopenic Obesity and Outcomes for Patients With Cancer. JAMA Netw. Open. Am. Med. Association . 7 (6), e2417115–e2417115. 10.1001/JAMANETWORKOPEN.2024.17115 (2024). Molwitz, I. et al. Muscle quality determined by computed tomography predicts short-term and long-term survival after liver transplantation. Sci. Rep. Sci. Rep. 13 (1). 10.1038/S41598-023-33349-Y (2023). Westenberg, L. B. et al. Reference values for low muscle mass and myosteatosis using tomographic muscle measurements in living kidney donors. Sci. Rep. Sci. Rep. 13 (1). 10.1038/S41598-023-33041-1 (2023). Sabatino, A. et al. Myosteatosis is associated with adiposity, metabolic derangements and mortality in patients with chronic kidney disease. Eur. J. Clin. Nutr. Eur. J. Clin. Nutr. 79 (5), 475–483. 10.1038/S41430-024-01551-4 (2025). Additional Declarations No competing interests reported. Supplementary Files FullyAutomatedCTBasedQuantitativeBodyCompositionAnalysisforPredictingSurvivalinPatientswithHCCUndergoingTACEADualCohortStudysupplementary.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 06 Apr, 2026 Reviews received at journal 02 Apr, 2026 Reviewers agreed at journal 28 Mar, 2026 Reviews received at journal 08 Feb, 2026 Reviewers agreed at journal 03 Feb, 2026 Reviewers invited by journal 03 Feb, 2026 Editor invited by journal 02 Feb, 2026 Editor assigned by journal 29 Jan, 2026 Submission checks completed at journal 29 Jan, 2026 First submitted to journal 27 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8713867","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":586532907,"identity":"5dd74920-630e-4005-a6ec-9259796a6e4a","order_by":0,"name":"Krzysztof Bartnik","email":"","orcid":"","institution":"Second Department of Radiology, Medical University of Warsaw","correspondingAuthor":false,"prefix":"","firstName":"Krzysztof","middleName":"","lastName":"Bartnik","suffix":""},{"id":586532909,"identity":"f5c599ec-891a-4de2-b586-9ede9c80cedd","order_by":1,"name":"Belgutei Tsevegmed","email":"","orcid":"","institution":"Second Department of Radiology, Medical University of Warsaw","correspondingAuthor":false,"prefix":"","firstName":"Belgutei","middleName":"","lastName":"Tsevegmed","suffix":""},{"id":586532911,"identity":"4ce55e06-e1e1-46a9-8af8-2a35e5021b29","order_by":2,"name":"Kamil Książek","email":"","orcid":"","institution":"Centre for Credible Artificial Intelligence, Warsaw University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Kamil","middleName":"","lastName":"Książek","suffix":""},{"id":586532914,"identity":"3413033a-e46d-4061-8567-628234aca9c1","order_by":3,"name":"Zuzanna Wojtczak","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABFklEQVRIiWNgGAWjYJCDBBBhwA/hMINwAyEtiSAVBpINcC2MBLWAVRgYHCCgRX5G+sPPhTvsGAyONzx/8OOPnbHxjRzTDQwV1okN0o1YtTDOSEiWnnkmmcHgzIHExh6eZDOzGzlmNxjOpCc2yBzEqoVZIuGANG8bM4PZjYTEZgYJZhuz20AtjG2HExskErFqYZNIbP7N21YP1WJQb2M8G6TlH24tPBLJbEBbDkO1JBw2M5AGaWnArUWC5xmbNW/bcR57oF9m9hw4bixx/1nZjYRj6cZtOPwi357++DZvW7WcZHtPwocff6oN+3sOb7vxocZatl+6+QA2LXAHAlECgpsA9iY+DWDAjm4mYS2jYBSMglEwMgAAadZjF4sBIQcAAAAASUVORK5CYII=","orcid":"","institution":"Second Department of Radiology, Medical University of Warsaw","correspondingAuthor":true,"prefix":"","firstName":"Zuzanna","middleName":"","lastName":"Wojtczak","suffix":""},{"id":586532916,"identity":"54e0a6cd-82a9-4755-aaaf-fef023b936d9","order_by":4,"name":"Przemysław Biecek","email":"","orcid":"","institution":"Centre for Credible Artificial Intelligence, Warsaw University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Przemysław","middleName":"","lastName":"Biecek","suffix":""}],"badges":[],"createdAt":"2026-01-27 19:38:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8713867/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8713867/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102181018,"identity":"bf76f632-a753-4366-8272-65d65c774246","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":379580,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the study. Two independent cohorts (WAW-TACE and HCC-TACE-Seg) were analyzed separately. Portal phase CT scans acquired prior to TACE were used for automated extraction of body composition metrics (muscle and fat), followed by sex-specific Z-score normalization. Feature selection was performed using clinical and body composition variables based on univariable Cox proportional hazards (CoxPH models and random survival forests (RSF). Selected features were then incorporated into multivariable survival models (CoxPH or RSF) for overall survival prediction, showing the prognostic relevance of muscle density, muscle area, and muscle volume.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/be0a86b61c0ad978ee147107.png"},{"id":102181019,"identity":"c4116ecf-b42b-4d9a-90b6-9f106986c843","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1077868,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative examples of body composition extremes on CT imaging: \u003cstrong\u003e(A)\u003c/strong\u003e a patient with high skeletal muscle mass and low adipose tissue, and \u003cstrong\u003e(B)\u003c/strong\u003e a patient with low skeletal muscle mass and high adipose tissue.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/0803a988cabac3164468ebce.png"},{"id":102181016,"identity":"766ca438-248b-4f13-926c-90b64a5231ac","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":481282,"visible":true,"origin":"","legend":"\u003cp\u003eKaplan-Meier curves for overall survival stratified by the binary indicator for skeletal muscle density at the L3 level in the WAW-TACE cohort. Patients were dichotomized into high and low muscle-density groups using an optimal cut-off derived from a Contal and O’Quigley-type maximally selected log-rank statistic. Patients with low skeletal muscle density are shown in red, whereas those with high skeletal muscle density are shown in blue. Numbers at risk are displayed below the plot.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/8d6e9aafcd8b1d5846f1e9ac.png"},{"id":102181017,"identity":"1689c0cf-9c28-4848-b616-3d33616c4992","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":342876,"visible":true,"origin":"","legend":"\u003cp\u003ePermutation feature importance analysis showing the mean decrease in concordance index for clinical variables (blue) and muscle-fat imaging features (red, shown in bold). Higher values indicate a greater contribution of the corresponding feature to model performance.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/c8b7238f2ebb31c770263703.png"},{"id":102181020,"identity":"b7cf96c4-9e59-401a-8ad9-29d9701a70e5","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":159645,"visible":true,"origin":"","legend":"\u003cp\u003eKaplan-Meier curves for overall survival stratified by the binary indicator for skeletal muscle density at the L3 level in the HCC-TACE-Seg cohort. Patients were dichotomized into high and low muscle density groups using an optimal cut-off derived from a Contal and O’Quigley-type maximally selected log-rank statistic. Patients with low skeletal muscle density are shown in red, whereas those with high skeletal muscle density are shown in blue. Numbers at risk are displayed below the plot.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/99cbb7c8bbdf89ef72b3e88c.png"},{"id":102297052,"identity":"b60869a5-6999-4c35-aa00-60995c63d83d","added_by":"auto","created_at":"2026-02-10 10:25:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3368071,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/170f88d0-892d-4da5-8594-6ad36d488478.pdf"},{"id":102181015,"identity":"3a819a7c-9631-4138-aca6-4f511237b832","added_by":"auto","created_at":"2026-02-09 07:15:53","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":50358,"visible":true,"origin":"","legend":"","description":"","filename":"FullyAutomatedCTBasedQuantitativeBodyCompositionAnalysisforPredictingSurvivalinPatientswithHCCUndergoingTACEADualCohortStudysupplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8713867/v1/c0e2086d070685aff6aec827.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Fully Automated CT-Based Quantitative Body Composition Analysis for Predicting Survival in Patients with HCC Undergoing TACE: A Dual-Cohort Study","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eTransarterial chemoembolization (TACE) is a first line, minimally invasive procedure performed primarily in patients with unresectable hepatocellular carcinoma (HCC) [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Most prognostic scores created for HCC mostly rely on laboratory evaluations of underlying liver disease combined with basic radiologic characteristics, such as the number of tumor lesions and their size [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Nevertheless, these models are unable to adequately represent the diversity of HCC patients and have limited predictive accuracy [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Even while two patients may present with tumors of similar size and morphology as well as comparable laboratory parameters, their entire physiological state and, thus, prognosis may differ significantly [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. As a result, there is a need for more individualized pre-treatment risk-stratification tools that can optimize treatment strategies.\u003c/p\u003e \u003cp\u003eIn recent years, the significance of individualized prognostication has been emphasized by a growing number of studies [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. This concept is reflected in the growing interest in frailty-based prognostication approaches [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Frailty is understood as reduced physiologic reserve and greater vulnerability to unfavorable treatment outcomes [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. In clinical practice, frailty is commonly assessed using subjective measures, such as the Clinical Frailty Scale (CFS), which, despite its widespread use, is limited by inter-observer variability [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. In turn, objective body composition analysis has recently been suggested as a scalable alternative for conventional frailty criteria like age or performance status [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Variables such as skeletal muscle volume, visceral fat area, and other morphometric indices have been proposed as significant indicators of outcomes in a variety of clinical contexts, including overall survival (OS) in patients with HCC or colorectal cancer [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Conventionally, the quantification of these factors relies on the laborious process of manually segmenting regions of interest, which requires hand-crafted input and potentially introduces bias [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Therefore, automated segmentation models have gained increasing attention. However, most of the studied models are not open access and are tailored to specific populations, which importantly restricts their wider clinical usage [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eA recently published model from Chen et al. enables standardized and fully automated segmentation and extraction of imaging features of muscle and adipose tissue volumes on computed tomography (CT) images, providing a comprehensive body composition profile [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. The pipeline was developed, standardized, and validated on a heterogeneous cohort of CT scans representing multiple patient populations, ensuring generalizability. Recent studies on colorectal cancer patients have demonstrated that features extracted using this method can be utilized for survival prediction and treatment-outcome assessment [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, the prognostic value of automatically derived body composition parameters in patients with HCC, particularly in the setting of TACE, remains insufficiently explored.\u003c/p\u003e \u003cp\u003eTherefore, the aim of our study was to evaluate whether body composition parameters derived using a standardized and automated approach may serve as useful prognostic markers for prediction of survival in patients with HCC treated with TACE. For this purpose, we applied a recently released an open-access access model for automated body composition analysis on CT imaging in two independent HCC cohorts.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003e The study was conducted in accordance with the Declaration of Helsinki. Institutional review board approval and informed consent were waived, as the study used fully anonymized, publicly available open-science datasets.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Patient cohorts\u003c/h2\u003e \u003cp\u003eIn this retrospective study, we utilized two independent cohorts: WAW-TACE and HCC-TACE-Seg. The WAW-TACE dataset includes 233 patients treated between 2016 and 2021 [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. All patients were treatment-naive and diagnosed with unresectable HCC, subsequently managed with TACE monotherapy. Baseline imaging consists of multiphase contrast-enhanced CT acquired within 90 days prior to the initial TACE cycle. In turn, the HCC-TACE-Seg dataset features 105 patients treated between 2002 and 2012 [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. All cases from this dataset represent confirmed HCC managed with TACE as the sole first-line therapy or as initial bridging therapy. Baseline imaging includes multiphase contrast-enhanced CT performed 1\u0026ndash;12 weeks (average 3 weeks) before the first TACE session.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Clinical variables\u003c/h2\u003e \u003cp\u003eThe clinical variables analyzed included the following:\u003c/p\u003e \u003cp\u003e1) For WAW-TACE dataset: age, sex, alpha-fetoprotein (AFP), total bilirubin, serum albumin, number of tumor lesions, tumor size, Child-Pugh score (CPS), serum creatinine, alanine aminotransferase (ALT), Barcelona Clinic Liver Cancer stage (BCLC), and international normalized ratio (INR). In addition, the following clinical scores were evaluated: ALBI score, mHAP II score, 6\u0026amp;12 score, and HAP score [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e2) For HCC-TACE-Seg dataset: sex, age, CPS, AFP, BCLC, and numeric clinical staging scores including CLIP (Cancer of the Liver Italian Program) score, Okuda, TNM (Tumor-Node-Metastasis).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Extraction of image-based body composition metrics\u003c/h2\u003e \u003cp\u003eAll segmentations and body composition feature extractions were performed using the portal venous phase of the pre-TACE CT examinations. For the extraction of body composition metrics, we utilized the open-source pretrained deep learning pipeline by Chen et al [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. The authors designed an automated, deep learning-based framework capable of segmenting skeletal muscle, visceral adipose tissue, and subcutaneous adipose tissue on axial CT scans of the abdomen and pelvis. The pipeline supports both 2D analysis at the L3 level and 3D volumetric assessment from T12 to L4, providing a comprehensive and standardized set of quantitative measurements, as described in the study by Hanxue et al [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe extracted body composition metrics included the following variables. For 2D measurements at the L3 level, these comprised skeletal muscle area (mm\u0026sup2;), skeletal muscle density (2D SMD) [HU], subcutaneous fat area (mm\u0026sup2;), visceral fat area (mm\u0026sup2;), intramuscular fat area (mm\u0026sup2;), total fat area (mm\u0026sup2;), and total body cross-sectional area (mm\u0026sup2;). For 3D measurements between T12 and L4, the extracted variables included skeletal muscle volume (mm\u0026sup3;), skeletal muscle density (3D SMD) [HU], subcutaneous fat volume (mm\u0026sup3;), visceral fat volume (mm\u0026sup3;), intramuscular fat volume (mm\u0026sup3;), total fat volume (mm\u0026sup3;), total body volume (mm\u0026sup3;), and craniocaudal body height between T12 and L4 (mm).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Outcome variables\u003c/h2\u003e \u003cp\u003eThe primary endpoint of the study was to evaluate the association between automated body composition variables and OS in HCC patients undergoing TACE. Secondary endpoints included 3-year mortality rates and the assessment of whether the addition of CT-derived body-profiling variables could significantly improve the predictive performance of state-of-the-art clinical scores, as well as the best-performing combination of clinical variables.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Statistical analysis\u003c/h2\u003e \u003cp\u003eQuantitative body composition metrics were standardized using sex-specific Z-score normalization to control for sex-related physiological differences that were suggested by previous studies [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. For each variable, sex-specific means and standard deviations were calculated, and normalized values were derived as Z-scores according to the formula: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Z=\\frac{x-{\\mu\\:}_{sex}}{{\\sigma\\:}_{sex}}\\)\u003c/span\u003e\u003c/span\u003e where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:x\\:\\)\u003c/span\u003e\u003c/span\u003e denotes the observed value, while \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mu\\:}_{sex}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\sigma\\:}_{sex}\\)\u003c/span\u003e\u003c/span\u003e represent the sex-specific mean and standard deviation, respectively. 2D SMD was dichotomized using a data-driven cut-off, resulting in a binary muscle density indicator (binary 2D SMD). The classification into low and high subgroups in the Kaplan-Meier analysis was performed using an optimal cut-off value, determined by maximizing the log-rank test statistic (Contal \u0026amp; O\u0026rsquo;Quigley-like method) [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBody-composition parameters and clinical variables were first evaluated using univariable Cox proportional hazards (CoxPH) analyses. Variables that reached statistical significance (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the univariable analysis were subsequently included in multivariable CoxPH models to assess their independent predictive value. The concordance index (C-index) was used to evaluate the predictive performance of the CoxPH models for OS. Additionally, the area under the ROC curve (AUC) was calculated for 3-year mortality. Performance differences between models were quantified using ΔC-index and ΔAUC(t), with bootstrap-based confidence intervals.\u003c/p\u003e \u003cp\u003eAdditionally, a machine learning-based approach was used to assess the impact of body composition parameters on OS prediction. Firstly, a 5-fold cross-validation (CV) experiment using Random Survival Forests (RSF) with all available variables and 100 estimators was performed. In each cross-validation step, a permutation-based feature importance was calculated. After five CV folds, variables having a positive mean importance score, i.e. causing a more accurate C-index estimation, were selected. Then, the RSF model with variables considered important in the previous step (both clinical and body composition) was compared with the model trained using only a significant subset of clinical variables. This evaluation was performed in another 5-fold cross-validation experimental scenario with 10 repetitions. Finally, to assess the significance of the difference in model performance, a t-test for related samples was performed.\u003c/p\u003e \u003cp\u003eAll statistical analyses and figure generation were performed in Python (Python Software Foundation 2024, Language Reference Version 3.13.0). Data handling was conducted with the \u003cem\u003epandas\u003c/em\u003e and \u003cem\u003eNumPy\u003c/em\u003e libraries. Survival analyses were performed using the \u003cem\u003elifelines\u003c/em\u003e and \u003cem\u003escikit-survival\u003c/em\u003e packages, including Cox proportional hazards regression (\u003cem\u003eCoxPHFitter\u003c/em\u003e) and calculation of Harrell\u0026rsquo;s concordance index. Standard Python utilities (\u003cem\u003eos\u003c/em\u003e, \u003cem\u003epathlib\u003c/em\u003e) and typing modules were used for file handling and code organization. Analyses were executed in a reproducible Python environment, and the complete source code and environment specifications can be provided upon reasonable request.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Characteristics of patient populations\u003c/h2\u003e \u003cp\u003eWAW-TACE dataset: the final analysis included 230 patients who underwent TACE and had an available portal venous phase. OS was 28.6 months (95% CI: 21.8\u0026ndash;33.8); 1-year and 3-year mortality rates were 26.5% and 59.7%, respectively.\u003c/p\u003e \u003cp\u003eHCC-TACE-Seg dataset: 100 patients who underwent TACE and had an available portal venous phase were included. OS was 24.0 months (95% CI: 19.6\u0026ndash;29.0). The 1-year and 3-year mortality rates were 22.1% and 73.6%, respectively.\u003c/p\u003e \u003cp\u003eA detailed characterization of both evaluated cohorts is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline clinical characteristics of patients included in the study in the WAW-TACE and HCC-TACE-Seg cohorts. Continuous variables are reported as medians and categorical variables as counts with percentages. Differences between cohorts were assessed using appropriate statistical tests for continuous and categorical variables, with corresponding \u003cem\u003ep\u003c/em\u003e values reported.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWAW-TACE (n\u0026thinsp;=\u0026thinsp;230)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHCC-TACE-Seg (n\u0026thinsp;=\u0026thinsp;100)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale sex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e48 (20.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e35 (35.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.01\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, median (years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e68.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.17\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChild-Pugh class, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.08\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e204 (88.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81 (81.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e26 (11.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e18 (18.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1 (1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAFP, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;400 ng/mL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e176 (76.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72 (72.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;\u0026thinsp;400 ng/mL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e54 (23.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28 (28.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAFP, median (ng/mL)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e32.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBCLC stage, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e168 (73.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15 (15.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e62 (27.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23 (23.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e62 (62.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 WAW-TACE\u003c/h2\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Univariable survival analysis\u003c/h2\u003e \u003cp\u003eFull univariate CoxPH results for all raw and sex-normalized fat and muscle-based predictors are provided in the Supplementary Tables S1 and S2, respectively. All subsequent muscle- and fat-related variables reported in the following sections refer to these sex-normalized measures. Among all image-derived variables, only 2D SMD and 3D SMD were significant predictors of OS. The corresponding hazard ratios were 0.84 (95% CI, 0.72\u0026ndash;0.98; p\u0026thinsp;=\u0026thinsp;0.029) for 2D SMD and 0.85 (95% CI, 0.73-1.00; p\u0026thinsp;=\u0026thinsp;0.045) for 3D SMD. Among clinical variables, serum albumin, tumor size, log-transformed AFP, total bilirubin, number of tumor lesions, and Child-Pugh score were significant predictors of OS (Supplementary Table S3). Among all variables evaluated in the univariable analysis, the mHAP-2 score was the strongest predictor of OS (HR 1.58, 95% CI 1.32\u0026ndash;1.89, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01) with the lowest C-index of 0.641 (Supplementary Table S3).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Determination of binary indicator for skeletal muscle density at the L3 level\u003c/h2\u003e \u003cp\u003eIn the cut-off selection procedure for 2D SMD, the optimal threshold was identified as -0.85 with the log-rank test p\u0026thinsp;\u0026lt;\u0026thinsp;0.001 (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). This binary SMD indicator was subsequently evaluated in univariate survival models. For OS, the binary SMD was significantly associated with mortality risk (HR\u0026thinsp;=\u0026thinsp;0.53, 95% CI: 0.37\u0026ndash;0.76, p\u0026thinsp;=\u0026thinsp;0.0005). For 3-year mortality, the association was also significant (HR\u0026thinsp;=\u0026thinsp;0.59, 95% CI: 0.40\u0026ndash;0.86, p\u0026thinsp;=\u0026thinsp;0.006). Patients below the cut-off had a median OS of 19.9 months (95% CI: 9.3\u0026ndash;32.8), whereas those above the cut-off had a median OS of 32.3 months (95% CI: 22.9\u0026ndash;38.4). At 3 years, survival rates were 22.8% (95% CI: 11.8\u0026ndash;36.0) and 44.5% (95% CI: 37.2\u0026ndash;51.5), respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3 Multivariable analysis\u003c/h2\u003e \u003cp\u003eAfter adjusting for the mHAP-2 score (the strongest clinical composite score associated with OS), binary 2D SMD remained independently associated with OS (HR 0.68, 95% CI 0.46-1.00, p\u0026thinsp;=\u0026thinsp;0.049). When evaluated as a predictive marker, adding binary 2D SMD to the mHAP-2 model resulted in only a small improvement in discrimination: the C-index increased from 0.641 to 0.657 (Δ\u0026thinsp;=\u0026thinsp;0.015; 95% CI -0.006 to 0.039; p\u0026thinsp;=\u0026thinsp;0.513). A similar modest gain was observed when adding continuous 2D SMD instead of the binary cut-off (C-index 0.657; Δ\u0026thinsp;=\u0026thinsp;0.016; 95% CI -0.001 to 0.033; p\u0026thinsp;=\u0026thinsp;0.485).\u003c/p\u003e \u003cp\u003eSimilar results were obtained in an analysis adjusting for individual clinical variables rather than composite clinical scores. Binary 2D SMD remained an independent predictor after adjustment for state-of-the-art clinical confounders identified in univariable analysis (Supplementary Table S4), including serum albumin, total bilirubin, log-transformed AFP, Child-Pugh score, tumor size, and number of tumor lesions. Incorporation of binary 2D SMD to clinical variables resulted in a modest improvement in model discrimination, with the concordance index increasing from 0.672 for the clinical model to 0.684 for the extended model (ΔC-index\u0026thinsp;=\u0026thinsp;0.011, 95% CI\u0026thinsp;\u0026minus;\u0026thinsp;0.003 to 0.027; bootstrap p\u0026thinsp;=\u0026thinsp;0.50).\u003c/p\u003e \u003cp\u003eA similar pattern was observed in time-specific discrimination. At 3 years, the time-dependent AUC increased from 0.698 for the clinical-only model to 0.720 after inclusion of binary 2D SMD (ΔAUC\u0026thinsp;=\u0026thinsp;0.022); however, this improvement did not reach statistical significance (95% CI, \u0026minus;\u0026thinsp;0.002 to 0.046; bootstrap p\u0026thinsp;=\u0026thinsp;0.50). A comparable improvement was observed when a continuous 2D SMD was used instead (ΔAUC\u0026thinsp;=\u0026thinsp;0.026, p\u0026thinsp;=\u0026thinsp;0.51).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e3.2.4 Machine learning-based feature importance estimation\u003c/h2\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe feature importance estimation with RSF revealed that 9 clinical and 3 muscle-based variables are valuable predictors with positive importance scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The most important variables were log-transformed AFP, total bilirubin, and tumor size, belonging to the category of clinical variables. The three selected muscle-based predictors, i.e. skeletal muscle area at the L3 level, 2D SMD, and skeletal muscle volume, were considered to have a positive influence on model estimations. The remaining 15 variables were excluded from further calculations.\u003c/p\u003e \u003cp\u003eIn the second cross-validation stage, a model with a subset of all variables was compared with a model having only a subset of clinical predictors. The combined model obtained the C-index of 0.6578, while the clinical-only model achieved 0.6352. Since the differences in predictions between the two models were normally distributed according to the results of the Shapiro-Wilk test (W\u0026thinsp;=\u0026thinsp;0.98, p\u0026thinsp;=\u0026thinsp;0.59), the one-sided t-test for related samples was calculated. Its results (t\u0026thinsp;=\u0026thinsp;6.04, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) indicate that the inclusion of muscle-based predictors together with clinical variables resulted in a statistically significant improvement in model performance compared with the clinical-only model.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.3 HCC-TACE-Seg\u003c/h2\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1 Univariable survival analysis\u003c/h2\u003e \u003cp\u003eFull univariate CoxPH results for all sex-normalized fat and muscle metrics are provided in the Supplementary Table S5. Among all evaluated factors, 2D SMD and 3D SMD were the only variables significantly associated with OS in the univariable analysis, which was analogous to results on the WAW-TACE dataset. The corresponding hazard ratios were 0.79 (95% CI, 0.65\u0026ndash;0.98; p\u0026thinsp;=\u0026thinsp;0.028) for 2D SMD and 0.80 (95% CI, 0.65\u0026ndash;0.99; p\u0026thinsp;=\u0026thinsp;0.045) for 3D SMD.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2 Determination of binary indicator for skeletal muscle density at the L3 level\u003c/h2\u003e \u003cp\u003eIn the cut-off selection procedure for 2D SMD, the optimal threshold was identified as 0.61, which yielded a statistically significant separation in survival curves (log-rank p\u0026thinsp;=\u0026thinsp;0.002). The dichotomized variable was subsequently evaluated in univariable CoxPH, demonstrating a strong association with overall survival (HR\u0026thinsp;=\u0026thinsp;0.43, 95% CI: 0.25\u0026ndash;0.75, p\u0026thinsp;=\u0026thinsp;0.0029). The median OS in the cohort was 24.0 months (95% CI: 19.6\u0026ndash;29.0). Patients classified as having low 2D SMD according to the binary indicator had a median OS of 24.0 months (95% CI: 17.5\u0026ndash;28.1), whereas those classified as having high 2D SMD had a median OS of 35.6 months (95% CI: 15.2\u0026ndash;81.6). At 3 years, OS was 18.6% (95% CI: 10.6\u0026ndash;28.5) and 48.0% (95% CI: 27.8\u0026ndash;65.6), respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3 Multivariable analysis\u003c/h2\u003e \u003cp\u003eWithin this dataset, the CLIP score emerged as the strongest clinical predictor of mortality (HR 1.43, 95% CI 1.15\u0026ndash;1.77, p\u0026thinsp;=\u0026thinsp;0.0012) (Supplementary Table S6). In the multivariable CoxPH model, including the CLIP score and binary 2D SMD, both covariates were significantly associated with OS. Binary 2D SMD showed a hazard ratio of 0.42 (95% CI: 0.24\u0026ndash;0.73, p\u0026thinsp;=\u0026thinsp;0.002).\u003c/p\u003e \u003cp\u003eIn the multivariable CoxPH analysis for OS, incorporation of the binary 2D SMD into the clinical model based on the CLIP score resulted in a numerically higher discriminative performance. Specifically, the C-index increased from 0.565 for the clinical-only model to 0.602 for the extended model including the binary 2D SMN, corresponding to an absolute ΔC-index of 0.037. However, this improvement did not reach statistical significance (95% CI, \u0026minus;\u0026thinsp;0.018 to 0.094; bootstrap p\u0026thinsp;=\u0026thinsp;0.498). At 3 years, time-dependent discrimination improved from an AUC of 0.579 to 0.704 (ΔAUC\u0026thinsp;=\u0026thinsp;0.125). While the bootstrap confidence interval suggested a non-zero effect (95% CI, 0.019\u0026ndash;0.234), bootstrap-based hypothesis testing did not indicate statistical significance (p\u0026thinsp;=\u0026thinsp;0.466).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we evaluated the prognostic significance of standardized CT-derived body composition parameters on survival in patients with unresectable HCC treated with TACE. The results were validated in two independent datasets, supporting their robustness and generalizability, which is a major strength of this study. Despite differences in sample size, timeframe, and patient characteristics across both cohorts, skeletal muscle-based metrics consistently were significant and independent predictors of OS. The reproducibility of this finding across two different clinical populations suggests that muscle quality captures important aspects of physiological reserve and patient frailty that traditional clinical scores do not fully capture. Specifically, in the WAW-TACE cohort, 2D SMD and 3D SMD demonstrated an association with survival, and the effect of 2D SMD remained preserved even after adjustment for established clinical scores such as the mHAP-2 score. Similarly, in the HCC-TACE-Seg dataset, both muscle density measurements (2D SMD and 3D SMD) were significantly associated with mortality risk, reinforcing the robustness of this metric across different clinical settings. At the same time, our evaluation of predictive performance showed that although muscle density remained independently associated with mortality in multivariable CoxPH models, its incremental contribution to overall model discrimination was modest. For example, in the WAW-TACE cohort, adding binary 2D SMD to a combination of significant clinical variables increased the C-index only slightly (ΔC-index\u0026thinsp;=\u0026thinsp;0.011), and a similar limited improvement was observed in the 3-year AUC (ΔAUC\u0026thinsp;=\u0026thinsp;0.022). Although the inclusion of muscle density consistently yielded numerically higher C-indices, the bootstrap CI crossed zero and the differences were not statistically significant, indicating that the observed improvements in discrimination were modest and likely reflect a small incremental prognostic value. These findings further suggest that the study may be underpowered to detect small but statistically significant differences in model performance. Importantly, subsequent machine learning evaluation further supported these findings, as demonstrated by the RSF analysis, in which a model incorporating a subset of both clinical and muscle-based variables achieved a significantly higher performance than the clinical-only model (ΔC-index\u0026thinsp;=\u0026thinsp;0.023; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003cp\u003eThese findings indicate that automated CT-based body composition analysis including muscle density assessment provides prognostically relevant information not fully captured by standard clinical variables, but the magnitude of improvement in model performance remains relatively small. In other words, automated skeletal muscle density may be valuable as a complementary imaging-derived biomarker within broader multimodal risk-stratification frameworks, rather than as a standalone predictor intended to substantially enhance discrimination. In particular, muscle density may be especially useful in machine learning-based and radiomics-driven approaches. Those models frequently rely on multiple imaging features, where SMD can serve as an additional biologically interpretable variable contributing incremental prognostic value and improving overall model performance [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Its integration into imaging-only models warrants further investigation.\u003c/p\u003e \u003cp\u003eOur results align with previous reports linking reduced muscle quality to poorer outcomes following colorectal surgery. A recent study by Hanxue Gu et al. showed that sex-normalized SMD -measured in the same way as in our study - was a strong and independent predictor of 1-year mortality after colectomy (OR 0.42) [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The authors showed that these metrics have prognostic value independent of traditional clinical variables and enhance the discriminatory accuracy of established risk calculators (ΔC-index\u0026thinsp;=\u0026thinsp;0.07, 0.80 vs. 0.73), whereas in our study, they improved prediction only modestly (ΔC-index values ranging from 0.011 to 0.037). Importantly, our similar observations in an entirely different clinical context support the idea that automated SMD represents a reliable biomarker that generalizes across a variety of clinical scenarios and imaging protocols. This implies that automated CT-based muscle quality assessment captures significant parts of patient frailty and physiologic reserve that are not represented by conventional clinical variables. As a result, these measures could be useful as adjuncts for individualized risk assessment in a variety of patient groups. This method is easy to implement and does not require additional procedures beyond standard clinical care. Once included in the clinical workflow, the model can be fully automated.\u003c/p\u003e \u003cp\u003eThese findings are consistent with a recent meta-analysis that summarizes the predictive significance of body composition in HCC patients. The pooled HR\u0026thinsp;=\u0026thinsp;1.38 for sarcopenia as a predictor of OS in TACE patients, suggesting that subjects with low muscle quality have a significantly higher mortality [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. However, the methodologies used to assess muscle mass and muscle quality differed substantially across the included studies, ranging from manual and semi-automated segmentations at different anatomical levels to threshold-based approaches and ROI measurements. This variability makes direct comparison more difficult, restricts wider clinical translation, and underscores the need for standardized, reproducible approaches to body composition evaluation. In this context, the fully automated and publicly available model used in our study offers a potential solution by providing consistent and observer-independent measurements that may facilitate standardization across centers and imaging protocols.\u003c/p\u003e \u003cp\u003eA conceptually similar automated approach was previously presented by Wang et al., who developed an in-house deep convolutional neural network to facilitate psoas muscle segmentation on CT studies [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Their study served primarily as a proof-of-concept that deep learning can replace manual measurements of psoas muscle volume and that such an automated psoas muscle index is independently associated with OS in patients with liver cirrhosis. A cohort of 254 people with chronic liver disease was used for clinical validation in their investigation. The study showed that automated psoas size measurement was a significant survival predictor even after adjusting for age, sex, and liver function. Specifically, the CT-derived psoas muscle index was significantly associated with survival (HR\u0026thinsp;=\u0026thinsp;0.86), but the study did not present parametric model performance metrics comparing models with and without this parameter. Moreover, the model used in that study was trained and assessed mostly on internal datasets, lacked external validation, and was not made available as a standardized, publicly accessible tool, that restricts its wider clinical applicability. Despite limitations, these results offered proof that prognostic information is carried by image-based muscle measures. Our study expands this concept beyond chronic liver disease in general to a more clinically specific context: TACE patients with HCC. Additionally, our analysis relied on an open-source, pretrained, fully automated pipeline that, once applied, does not require manual annotation or additional radiologist input, unlike Wang et al.'s approach. This substantially lowers the barrier to integration in other clinical centers and supports the feasibility of using standardized body-composition profiling as part of routine pre-TACE risk stratification.\u003c/p\u003e \u003cp\u003eA series of related observations was reported by Surov et al., who studied the prognostic role of muscle quality in patients with HCC treated with different locoregional therapies. In one study, involving patients undergoing combined SIRT and sorafenib treatment, skeletal muscle density was measured semi-automatically at the L3 level with manual radiologist corrections [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Myosteatosis was associated with OS in the univariable analysis, although this effect did not remain significant after adjustment for demographic and clinical variables. Notably, subgroup analyses showed that myosteatosis was a significant predictor only in patients with alcohol-related cirrhosis, suggesting that its importance may depend on underlying liver disease. In another study by the same authors, using a similar semi-automated methodology in cirrhotic patients treated with TACE, myosteatosis again did not demonstrate an independent association with survival, whereas the albumin-muscle density score emerged as a significant predictor [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Both studies reported associations with survival but did not include a formal evaluation of the additive model performance of imaging-derived muscle parameters beyond hazard ratios.\u003c/p\u003e \u003cp\u003eThe broader prognostic relevance of body composition measurements has been demonstrated across multiple clinical settings. Sarcopenia and myosteatosis have been associated with increased mortality in patients with breast cancer, chronic liver disease, chronic kidney disease, and liver transplantation [\u003cspan additionalcitationids=\"CR31 CR32 CR33\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Collectively, these results suggest that muscle mass and muscle quality are significant indicators of physiological susceptibility in oncology and other medical scenarios. This supports the idea that a fully automated, standardized CT-based methodology, like the one employed in our study, could be potentially applicable in a variety of illness scenarios that require individualized risk stratification.\u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Limitations\u003c/h2\u003e \u003cp\u003eThis study has several limitations that should be acknowledged. First, the retrospective design of both cohorts introduces the possibility of selection bias and unmeasured confounding that may influence outcomes independently of body composition metrics. Second, although the use of two independent datasets enhances the generalizability of our findings, the cohorts differed with respect to treatment timeframe (2002\u0026ndash;2012 vs. 2016\u0026ndash;2021), clinical practice standards, imaging protocols, scanner types, and baseline patient characteristics. These sources of heterogeneity, combined with the lack of scanner-specific normalization, may affect the comparability of absolute muscle-density values and the magnitude of observed associations. Finally, the optimal cut-off for muscle density was derived using maximally selected log-rank statistics, which carries an inherent risk of overfitting. Data-driven thresholds may overestimate effect sizes unless validated prospectively or in an independent test set.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eIn patients with unresectable HCC treated with TACE, automated, quantitative CT-based assessment of skeletal muscle density emerged as an independent predictor of OS. This effect was reproducible in two independent cohorts and consistent with findings from an external colorectal cancer study, suggesting broader relevance beyond HCC. Although its addition led to only modest improvements in discrimination over established clinical models, the automated muscle-density measure is easy to implement, scalable, and may serve as a useful complementary quantitative imaging biomarker within multimodal risk-stratification frameworks.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis project was supported by the SONATINA grant from the National Science Centre (NCN), No. 2025/56/C/NZ7/00276 (DEC-2025/56/C/NZ7/00276).\u003c/p\u003e\n\u003ch2\u003eAuthor contributions statement\u003c/h2\u003e\n\u003cp\u003eStudy design: KB; KK\u003c/p\u003e\n\u003cp\u003eMethodology: KB, KK\u003c/p\u003e\n\u003cp\u003eStatistical analysis: KB, KK, ZW\u003c/p\u003e\n\u003cp\u003eSegmentation and feature extraction: BT\u003c/p\u003e\n\u003cp\u003eVisualization: KB, KK, ZW\u003c/p\u003e\n\u003cp\u003eSupervision: PB\u003c/p\u003e\n\u003cp\u003eWriting\u0026ndash;Original Draft Preparation: KB\u003c/p\u003e\n\u003cp\u003eWriting\u0026ndash;Review, Feedback \u0026amp; Editing: all authors\u003c/p\u003e\n\u003cp\u003eApproval of final version of submitted manuscript: all authors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003ch2\u003e\u0026nbsp;\u003c/h2\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eStudy design: KB, KK; Methodology: KB, KK; Statistical analysis: KB, KK, ZW; Segmentation and feature extraction: BT; Visualization: KB, KK, ZW; Supervision: PB; Writing\u0026ndash;Original Draft Preparation: KB; Writing\u0026ndash;Review, Feedback \u0026amp; Editing: all authors; Approval of final version of submitted manuscript: all authors.\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eAll datasets used in this study are publicly available. The WAW-TACE dataset is available in the Radiology: Artificial Intelligence under DOI: 10.1148/RYAI.240296. The HCC‑TACE‑Seg dataset is available in the Scientific Data under DOI: 10.1038/s41597‑023‑01928‑3.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMarrero, J. A. et al. Diagnosis, Staging, and Management of Hepatocellular Carcinoma: 2018 Practice Guidance by the American Association for the Study of Liver Diseases. \u003cem\u003eHepatol. Hepatol.\u003c/em\u003e \u003cb\u003e68\u003c/b\u003e (2), 723\u0026ndash;750. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/HEP.29913\u003c/span\u003e\u003cspan address=\"10.1002/HEP.29913\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeimbach, J. K. et al. \u003cem\u003eAASLD guidelines for the treatment of hepatocellular carcinoma. Hepatology\u003c/em\u003e Vol. 67, 358\u0026ndash;380 (John Wiley and Sons Inc., 2018). 1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, Q. et al. Development of a prognostic score for recommended TACE candidates with hepatocellular carcinoma: A multicentre observational study. \u003cem\u003eJ. Hepatol. J. Hepatol.\u003c/em\u003e \u003cb\u003e70\u003c/b\u003e (5), 893\u0026ndash;903. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/J.JHEP.2019.01.013\u003c/span\u003e\u003cspan address=\"10.1016/J.JHEP.2019.01.013\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson, P. J. et al. Assessment of liver function in patients with hepatocellular carcinoma: A new evidence-based approach - The albi grade. \u003cem\u003eJ. Clin. Oncol. Am. Soc. Clin. Oncol.\u003c/em\u003e \u003cb\u003e33\u003c/b\u003e (6), 550\u0026ndash;558. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1200/JCO.2014.57.9151\u003c/span\u003e\u003cspan address=\"10.1200/JCO.2014.57.9151\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePinato, D. J. et al. Combined sequential use of HAP and ART scores to predict survival outcome and treatment failure following chemoembolization in hepatocellular carcinoma: A multi-center comparative study. \u003cem\u003eOncotarget Impact Journals LLC\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e (28), 44705\u0026ndash;44718. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.18632/oncotarget.9604\u003c/span\u003e\u003cspan address=\"10.18632/oncotarget.9604\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKelley, R. K. \u0026amp; Greten, T. F. Hepatocellular Carcinoma - Origins and Outcomes. New England Journal of Medicine. \u003cem\u003eMass. Med. Soc.\u003c/em\u003e \u003cb\u003e385\u003c/b\u003e (3), 280\u0026ndash;282 (2021). doi: 10.1056/NEJMCIBR2106594;ISSUE:ISSUE:DOI.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVitale, A., Trevisani, F., Farinati, F. \u0026amp; Cillo, U. Treatment of Hepatocellular Carcinoma in the Precision Medicine Era: From Treatment Stage Migration to Therapeutic Hierarchy. Hepatology: Official Journal of the American Association for the Study of Liver Diseases. John Wiley and Sons Inc; ;72(6):2206\u0026ndash;2218. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/HEP.31187\u003c/span\u003e\u003cspan address=\"10.1002/HEP.31187\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVitale, A. et al. Personalised management of patients with hepatocellular carcinoma: a multiparametric therapeutic hierarchy concept. \u003cem\u003eLancet Oncol. Elsevier Ltd.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e (7), e312\u0026ndash;e322. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S1470-2045(23)00186-9\u003c/span\u003e\u003cspan address=\"10.1016/S1470-2045(23)00186-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHamaker, M. E. et al. Frailty screening methods for predicting outcome of a comprehensive geriatric assessment in elderly patients with cancer: A systematic review. \u003cem\u003eLancet Oncol. Elsevier\u003c/em\u003e. \u003cb\u003e13\u003c/b\u003e (10), e437\u0026ndash;e444. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S1470-2045(12)70259-0\u003c/span\u003e\u003cspan address=\"10.1016/S1470-2045(12)70259-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEthun, C. G. et al. Frailty and cancer: Implications for oncology surgery, medical oncology, and radiation oncology. \u003cem\u003eCA Cancer J. Clin. Wiley\u003c/em\u003e. \u003cb\u003e67\u003c/b\u003e (5), 362\u0026ndash;377 (2017). .21406;REQUESTEDJOURNAL:JOURNAL:15424863;WGROUP:STRING:PUBLICATION.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWelford, J. et al. The Clinical Frailty Scale can indicate prognosis and care requirements on discharge in oncology and haemato-oncology inpatients: A cohort study. \u003cem\u003eEur. J. Cancer Care (Engl) John Wiley Sons Inc\u003c/em\u003e. \u003cb\u003e31\u003c/b\u003e (6), e13752. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/ECC.13752\u003c/span\u003e\u003cspan address=\"10.1111/ECC.13752\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShah, D. et al. Use of the Rockwood Clinical Frailty Scale in patients with advanced hepatopancreaticobiliary malignancies. \u003cem\u003eExpert Rev. Anticancer Ther. Expert Rev. Anticancer Ther.\u003c/em\u003e \u003cb\u003e22\u003c/b\u003e (9), 1009\u0026ndash;1015. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/14737140.2022.2096594\u003c/span\u003e\u003cspan address=\"10.1080/14737140.2022.2096594\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, Y. et al. Automated Muscle and Fat Segmentation in Computed Tomography for Comprehensive Body Composition Analysis. Machine Learning for Biomedical Imaging. \u003cem\u003eMach. Learn. Biomedical Imaging\u003c/em\u003e. \u003cb\u003e3\u003c/b\u003e (November 2025), 581\u0026ndash;617. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.59275/J.MELBA.2025-AG5G\u003c/span\u003e\u003cspan address=\"10.59275/J.MELBA.2025-AG5G\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu, H. et al. Improving Surgical Risk Prediction Through Integrating Automated Body Composition Analysis: a Retrospective Trial on Colectomy Surgery. ; (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/pdf/2506.11996\u003c/span\u003e\u003cspan address=\"https://arxiv.org/pdf/2506.11996\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed December 11, 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, A. et al. \u003cem\u003eImpact of body composition on the prognosis of hepatocellular carcinoma patients treated with transarterial chemoembolization: A systematic review and meta-analysis. Heliyon\u003c/em\u003e Vol. 10, e25237 (Elsevier Ltd, 2024). 310.1016/J.HELIYON.2024.E25237\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, N. C. et al. Automated Measurements of Muscle Mass Using Deep Learning can Predict Clinical Outcomes in Patients with Liver Disease. \u003cem\u003eAm. J. Gastroenterol. Wolters Kluwer Health\u003c/em\u003e. \u003cb\u003e115\u003c/b\u003e (8), 1210. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.14309/AJG.0000000000000662\u003c/span\u003e\u003cspan address=\"10.14309/AJG.0000000000000662\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGitHub - mazurowski-lab/CT-Muscle-and-Fat-Segmentation. (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/mazurowski-lab/CT-Muscle-and-Fat-Segmentation\u003c/span\u003e\u003cspan address=\"https://github.com/mazurowski-lab/CT-Muscle-and-Fat-Segmentation\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed December 11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBartnik, K. et al. WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data. \u003cem\u003eRadiol. Artif. Intell. Radiological Soc. North. Am. Inc\u003c/em\u003e. \u003cb\u003e6\u003c/b\u003e (6). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1148/RYAI.240296/ASSET\u003c/span\u003e\u003cspan address=\"10.1148/RYAI.240296/ASSET\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e/IMAGES/LARGE/RYAI.240296.FIG3.JPEG\u003c/span\u003e\u003cspan address=\"http:///IMAGES/LARGE/RYAI.240296.FIG3.JPEG\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoawad, A. W. et al. Multimodality annotated hepatocellular carcinoma data set including pre- and post-TACE with imaging segmentation. \u003cem\u003eSci. Data Nat. Publishing Group.\u003c/em\u003e ; ;\u003cb\u003e10\u003c/b\u003e(1). doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/S41597-023-01928-3\u003c/span\u003e\u003cspan address=\"10.1038/S41597-023-01928-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKadalayil, L. et al. A simple prognostic scoring system for patients receiving transarterial embolisation for hepatocellular cancer. Annals of Oncology. \u003cem\u003eAnn. Oncol.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e (10), 2565\u0026ndash;2570. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/annonc/mdt247\u003c/span\u003e\u003cspan address=\"10.1093/annonc/mdt247\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark, Y. et al. Addition of tumor multiplicity improves the prognostic performance of the hepatoma arterial-embolization prognostic score. \u003cem\u003eLiver Int. Liver Int.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e (1), 100\u0026ndash;107. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/LIV.12878\u003c/span\u003e\u003cspan address=\"10.1111/LIV.12878\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee, I. C. et al. A new ALBI-based model to predict survival after transarterial chemoembolization for BCLC stage B hepatocellular carcinoma. \u003cem\u003eLiver Int. Liver Int.\u003c/em\u003e \u003cb\u003e39\u003c/b\u003e (9), 1704\u0026ndash;1712. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/LIV.14194\u003c/span\u003e\u003cspan address=\"10.1111/LIV.14194\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDaniele, B. et al. \u003cem\u003eCancer of the Liver Italian Program (CLIP) score for staging hepatocellular carcinoma\u003c/em\u003e Vol. 37, S206\u0026ndash;S209 (Hepatology Research. John Wiley \u0026amp; Sons, Ltd, 2007). s210.1111/J.1872-034X.2007.00186.X\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKinoshita, A. et al. Staging systems for hepatocellular carcinoma: Current status and future perspectives. \u003cem\u003eWorld J. Hepatol. Baishideng Publishing Group. Co.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e (3), 406. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4254/WJH.V7.I3.406\u003c/span\u003e\u003cspan address=\"10.4254/WJH.V7.I3.406\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeister, F. A. et al. The Role of Sarcopenia and Myosteatosis in Short- and Long-Term Outcomes Following Curative-Intent Surgery for Hepatocellular Carcinoma in a European Cohort. \u003cem\u003eCancers (Basel) Cancers (Basel)\u003c/em\u003e. \u003cb\u003e14\u003c/b\u003e (3). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/CANCERS14030720\u003c/span\u003e\u003cspan address=\"10.3390/CANCERS14030720\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eContal, C. \u0026amp; O\u0026rsquo;Quigley, J. An application of changepoint methods in studying the effect of age on survival in breast cancer. \u003cem\u003eComput. Stat. Data Anal. N.-Holl.\u003c/em\u003e \u003cb\u003e30\u003c/b\u003e (3), 253\u0026ndash;270. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S0167-9473(98)00096-6\u003c/span\u003e\u003cspan address=\"10.1016/S0167-9473(98)00096-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (1999).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBartnik, K. et al. A novel radiomics approach for predicting TACE outcomes in hepatocellular carcinoma patients using deep learning for multi-organ segmentation. Scientific Reports Nature Publishing Group; 2024;14(1):1\u0026ndash;13. (2024). 14:1 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-024-65630-z\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-65630-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSurov, A. et al. Skeletal muscle quality predicts overall survival in advanced liver hepatocellular carcinoma treated with SIRT and sorafenib: A subanalysis of the SORAMIC trial. \u003cem\u003eUnited Eur. Gastroenterol. J. United Eur. Gastroenterol. J.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (8), 1016\u0026ndash;1027. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/UEG2.12627\u003c/span\u003e\u003cspan address=\"10.1002/UEG2.12627\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSurov, A. et al. Albumin-muscle density score predicts overall survival in patients with hepatocellular cancer undergoing treatment with transarterial chemoembolization. \u003cem\u003eJ. Cancer Res. Clin. Oncol. J. Cancer Res. Clin. Oncol.\u003c/em\u003e \u003cb\u003e150\u003c/b\u003e (12). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/S00432-024-06043-3\u003c/span\u003e\u003cspan address=\"10.1007/S00432-024-06043-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGourd, E. Sarcopenia and adiposity linked to overall survival. \u003cem\u003eLancet Oncol. NLM (Medline)\u003c/em\u003e. \u003cb\u003e19\u003c/b\u003e (5), e239. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/S1470-2045(18)30284-5\u003c/span\u003e\u003cspan address=\"10.1016/S1470-2045(18)30284-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, C. et al. Sarcopenic Obesity and Outcomes for Patients With Cancer. \u003cem\u003eJAMA Netw. Open. Am. Med. Association\u003c/em\u003e. \u003cb\u003e7\u003c/b\u003e (6), e2417115\u0026ndash;e2417115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1001/JAMANETWORKOPEN.2024.17115\u003c/span\u003e\u003cspan address=\"10.1001/JAMANETWORKOPEN.2024.17115\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMolwitz, I. et al. Muscle quality determined by computed tomography predicts short-term and long-term survival after liver transplantation. \u003cem\u003eSci. Rep. Sci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/S41598-023-33349-Y\u003c/span\u003e\u003cspan address=\"10.1038/S41598-023-33349-Y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWestenberg, L. B. et al. Reference values for low muscle mass and myosteatosis using tomographic muscle measurements in living kidney donors. \u003cem\u003eSci. Rep. Sci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/S41598-023-33041-1\u003c/span\u003e\u003cspan address=\"10.1038/S41598-023-33041-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSabatino, A. et al. Myosteatosis is associated with adiposity, metabolic derangements and mortality in patients with chronic kidney disease. \u003cem\u003eEur. J. Clin. Nutr. Eur. J. Clin. Nutr.\u003c/em\u003e \u003cb\u003e79\u003c/b\u003e (5), 475\u0026ndash;483. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/S41430-024-01551-4\u003c/span\u003e\u003cspan address=\"10.1038/S41430-024-01551-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2025).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8713867/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8713867/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTransarterial chemoembolization (TACE) is a standard treatment for patients with unresectable hepatocellular carcinoma (HCC), yet existing models provide limited individualized risk stratification. Automated CT-derived body composition analysis has emerged as an objective marker of patient physiological reserve, but its value in prognostication in TACE patients is insufficiently studied. Therefore, the aim of the study was to evaluate the prognostic value of a fully automated, open-source pipeline for CT-based body composition analysis in predicting overall survival (OS) in patients with HCC undergoing TACE. In this study, we used two independent cohorts of treatment-naive patients undergoing TACE: the WAW-TACE cohort (development) (n\u0026thinsp;=\u0026thinsp;230) and the HCC-TACE-Seg cohort (validation) (n\u0026thinsp;=\u0026thinsp;100). Skeletal muscle and fat metrics were extracted from pre-treatment CTs using a standardized deep learning pipeline and normalized by sex. Survival analyses were performed using Cox proportional hazards (CoxPH) models and random survival forests (RSF). Skeletal muscle density (SMD) at the L3 level was the strongest and independent predictor of OS across both cohorts (HR: development, 0.84; p\u0026thinsp;=\u0026thinsp;0.029; validation, 0.79; p\u0026thinsp;=\u0026thinsp;0.028). This association remained significant after adjustment for the best-performing clinical composite scores: mHAP-2 in the development (adjusted HR\u0026thinsp;=\u0026thinsp;0.68; p\u0026thinsp;=\u0026thinsp;0.049) and CLIP in the validation cohort (adjusted HR\u0026thinsp;=\u0026thinsp;0.43; p\u0026thinsp;=\u0026thinsp;0.003). In CoxPH, the addition of SMD metrics resulted in only modest improvements in discrimination (ΔC-index 0.011\u0026ndash;0.037) that did not reach statistical significance. In contrast, RSF analysis demonstrated a statistically significant improvement in model discrimination when muscle-based variables were added to clinical features (ΔC-index\u0026thinsp;=\u0026thinsp;0.023; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Fully automated CT-derived body composition metrics were independently associated with OS in patients treated with TACE. While their addition to traditional clinical models resulted in only modest improvements in discrimination, muscle density provided complementary prognostic information. Notably, statistically significant increases in model performance were observed when muscle-based features were integrated into RSF, highlighting their potential value within machine learning frameworks.\u003c/p\u003e \u003cp\u003eTable of Contents\u003c/p\u003e","manuscriptTitle":"Fully Automated CT-Based Quantitative Body Composition Analysis for Predicting Survival in Patients with HCC Undergoing TACE: A Dual-Cohort Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-09 07:15:48","doi":"10.21203/rs.3.rs-8713867/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-06T06:04:22+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-02T18:15:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"189025915080824484470794944138544842272","date":"2026-03-28T14:58:39+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-08T23:13:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"140263791881226201372848853553493227704","date":"2026-02-04T00:42:29+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-03T15:38:12+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-02T16:59:09+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-29T12:01:45+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-29T12:00:41+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-01-27T19:28:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e9bdb0bc-4f9a-41ef-8bc5-75938a72b143","owner":[],"postedDate":"February 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":62405934,"name":"Health sciences/Biomarkers"},{"id":62405935,"name":"Biological sciences/Cancer"},{"id":62405936,"name":"Health sciences/Medical research"},{"id":62405937,"name":"Health sciences/Oncology"},{"id":62405938,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2026-05-15T04:54:30+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-09 07:15:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8713867","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8713867","identity":"rs-8713867","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}