Abstract
Purpose
To evaluate whether whole-body PET/CT-derived body composition features are associated with survival in patients with resectable non-small cell lung cancer (NSCLC), using double machine learning to quantify adjusted associations with restricted mean survival time.
Methods
This retrospective multicenter study included 769 patients with stage ≤ IIIA NSCLC who underwent preoperative 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) and curative resection. Center 1 was used for model development (n = 555) and Center 2 for external validation (n = 214). Automated whole-body segmentation was used to extract volumetric, attenuation-based and metabolic features of skeletal muscle and adipose tissue. Restricted mean survival time-based double machine learning with generalized propensity score weighting estimated DML-adjusted differences in restricted mean survival time for overall survival (OS) and progression-free survival (PFS). Ridge-penalized Cox models were further used to evaluate prognostic performance and incremental value.
Results
Higher intermuscular adipose tissue (IMAT) volume index was associated with shorter survival (DML-adjusted RMST difference, -4.29 months for OS and − 2.74 months for PFS per 1-SD higher IMAT volume index). Higher TAT SUR (Mean) showed an exploratory favorable adjusted association with longer OS (+ 4.42 months). Sex-stratified analyses suggested stronger adverse adipose-volume associations in male patients and stronger favorable adipose-metabolic associations in female patients. Model analyses showed moderate external discrimination, whereas incremental-value metrics were modest and endpoint-dependent.
Conclusion
Whole-body PET/CT body-composition phenotyping may provide prognostic information in resectable NSCLC. Higher IMAT burden was associated with shorter survival, whereas TAT SUR (Mean) showed an exploratory favorable adjusted association with OS. A local software framework may support reproducible feature extraction and research-oriented risk stratification.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/s00259-026-08098-5.
Keywords: Non-small cell lung cancer, Body composition, PET/CT, Machine learning, Adjusted association, Myosteatosis
Introduction
Lung cancer remains the leading cause of cancer-related mortality worldwide, despite substantial advances in surgical techniques, systemic therapies, and perioperative management [1]. For patients with early-stage or resectable non-small cell lung cancer (NSCLC), anatomical lung resection offers the best chance for long-term survival; however, postoperative outcomes vary considerably even among patients with similar tumor stage and treatment strategies [2, 3]. This heterogeneity highlights the critical role of host-related factors beyond traditional clinicopathological variables.
Traditionally, body composition research has focused on assessing localized tissue depletion via single-slice computed tomography (CT) images, typically at the third lumbar (L3) or twelfth thoracic (T12) vertebral level, under the assumption that these single cross-sections accurately reflect total body mass [4, 5]. Although single-slice muscle and adipose tissue cross-sectional areas correlate with baseline wasting, they fail to capture the holistic, three-dimensional (3D) spatial distribution and volumetric heterogeneity of complex body composition compartments [6]. Furthermore, structural imaging alone remains blind to the varying metabolic states of these tissues. For instance, low-attenuation muscle or pathological adipose accumulation may remain structurally undetectable while driving silent, systemic chronic inflammation and altered host immunity, which profoundly influence the prognosis of non-small cell lung cancer (NSCLC) [7, 8].
Recent advances in deep learning-based segmentation frameworks, such as TotalSegmentator, have catalyzed a shift toward fully automated, reproducible 3D volumetric profiling of body composition across comprehensive anatomic regions [9]. Concurrently, integrating positron emission tomography with fluorodeoxyglucose (18F-FDG PET) expands this structural assessment by capturing tissue-specific metabolic activity, offering deeper insights into systematic energy expenditure and host inflammatory status [10, 11]. However, translating these multidimensional, multi-modal imaging phenotypes into reliable clinical prognostications presents a significant methodological bottleneck [12]. While utilizing full-volume, multi-modal imaging phenotypes captures a more holistic picture of host-tumor crosstalk, it simultaneously floods the analysis with highly correlated, high-dimensional covariates [12, 13]. In retrospective cohort studies, these interrelated imaging variables are heavily confounded by baseline clinical characteristics and selection biases [13]. Consequently, traditional regression frameworks—such as standard Cox proportional hazards models—are frequently inadequate to isolate the true, unconfounded independent prognostic effects of specific tissue compartments, rendering them highly susceptible to over-fitting and biased estimations [12, 14].
Despite growing interest, most prior studies evaluating body composition in lung cancer rely on single-slice CT metrics and conventional regression models, which may be limited in handling high-dimensional confounding and in estimating adjusted survival associations from correlated imaging phenotypes. These limitations are particularly relevant when analyzing imaging-derived phenotypes that are strongly correlated with age, sex, body mass index (BMI), comorbidities, and tumor burden. Double machine learning (DML) offers a causal-inference-inspired framework for quantifying adjusted associations under explicit assumptions, including adequate measurement of relevant confounders and sufficient overlap across exposure distributions, by flexibly modeling nuisance functions using machine learning while preserving valid statistical inference [12–14]. When combined with survival endpoints using restricted mean survival time (RMST), DML offers a robust and clinically interpretable alternative to hazard ratio-based analyses [13, 15].
Therefore, the aim of this study was to quantify the adjusted associations of whole-body PET/CT-derived body composition phenotypes with OS and PFS in patients undergoing curative-intent lung resection. Specifically, we evaluated volumetric, attenuation-based, and metabolic tissue features using a double machine learning framework and evaluated their incremental prognostic value over standard clinical variables.
Materials and methods
Study design
This retrospective study was approved by the institutional review board (2024SL044), which covered both participating centers, and informed consent was waived due to its retrospective design. Patients with pathologically confirmed NSCLC who underwent preoperative 18F-FDG PET/CT were retrospectively enrolled from 2 centers in our retrospective analysis. Center 1 (Second Affiliated Hospital of Naval Medical University) enrolled patients between January 2016 and March 2023, and Center 2 (Huadong Hospital) between July 2018 and December 2023. The inclusion criteria were as follows: (a) pathologically confirmed NSCLC of stage IIIA or lower treated with wedge resection, segmentectomy, or lobectomy; (b) standard 18F-FDG PET/CT conducted within one month prior to surgery; (c) patients aged over 18 years; (d) complete clinical and imaging data. The exclusion criteria included: (a) prior therapy (radiotherapy, chemotherapy, or targeted therapy) before 18F-FDG PET/CT; (b) patients with other cancer types; (c) missing follow-up data. The study design and analysis workflow are illustrated in (Fig. 1A).
Fig. 1.

Study design and analytic workflow illustrating patient selection, automated whole-body segmentation, DML-adjusted RMST association analysis using RMST-based double machine learning, and survival model construction (A). Flowchart of patient selection (B)
Clinical data collection and endpoints
Clinical variables were extracted from electronic medical records and included age, sex, smoking history, pathological subtype, lesion location, size and density, clinical stage, surgical procedure, and postoperative treatment. The primary endpoint was overall survival (OS), defined as the time from surgery to death from any cause or last follow-up. The secondary endpoint was progression-free survival (PFS), defined as the time from surgery to disease progression or death, whichever occurred first. Patients were followed every 3–6 months through outpatient visits or telephone interviews, with follow-up censored on September 30, 2025.
PET/CT acquisition
At both centers, all patients fasted for at least 6 h prior to the PET/CT scan. 18F-FDG was administered according to the patient’s weight, following the standard of 0.10–0.15 mCi/kg. In Center 1, PET/CT was performed using a Siemens Biograph 64 camera. The CT attenuation correction acquisition parameters were a voltage of 120 kV, a tube current of 256 mA, and a slice thickness of 3 mm. PET scans were acquired using 5–7 bed positions, with each position scanned for 2 min, and were reconstructed with a 512 × 512 matrix and a voxel size of 3 × 3 × 3 mm³. In Center 2, PET/CT was carried out using a Philips Vereos scanner. The CT acquisition parameters were a voltage of 120 kV, a tube current of 200 mA, and a slice thickness of 4 mm. PET scans were obtained using 6–8 bed positions, with each position scanned for 1.5 min, and then were reconstructed using a 512 × 512 matrix and a voxel size of 4 × 4 × 4 mm³. The images from the two centers were pre-processed with normalization and resampling before segmentation.
Automated whole-body composition segmentation
Whole-body tissue segmentation was performed on CT images using TotalSegmentator version 2.13.0 (https://github.com/wasserth/TotalSegmentator), a validated deep learning framework based on the nnU-Net architecture [9]. The tissue_4_types task was used to segment four tissue classes: skeletal_muscle, subcutaneous_fat, intermuscular_fat, and torso_fat. In this manuscript, these labels are reported as skeletal muscle (SM), subcutaneous adipose tissue (SAT), intermuscular adipose tissue (IMAT), and torso adipose tissue (TAT), respectively. TAT was defined as total torso fat within the thoracic, abdominal, and pelvic cavities, excluding subcutaneous fat. This anatomical definition of the torso_fat label has been reviewed and validated by two senior nuclear medicine physicians, and differs from conventional VAT masks. Bone structures used for normalization were obtained from the corresponding TotalSegmentator anatomical labels. All segmentations were visually reviewed to identify gross segmentation errors (Fig. 2).
Fig. 2.

Illustration of whole-body composition analysis, showing segmentation and feature-extraction results for skeletal muscle, subcutaneous adipose tissue, intermuscular adipose tissue, and torso adipose tissue
Feature extraction and normalization
For each tissue compartment, CT-derived volumetric and attenuation-based features were extracted, including volume, mean attenuation (HU), and the standard deviation of attenuation values, hereafter referred to as density (Mean) and density (SD), respectively. Due to variability in cranio-caudal scan coverage inherent to whole-body PET/CT, tissue volumes of SM, SAT, and IMAT were normalized to corresponding bone volumes, which served as a stable internal anatomical reference [16]. TAT volumes, which were fully covered in all scans, were normalized by patient height squared (m2) to generate standardized indices. Volume-based ratios were also calculated, including the IMAT-to-SM volume ratio and the TAT-to-SAT volume ratio.
PET-derived metabolic features included the mean and standard deviation of voxel-wise standardized uptake values (SUVs) within each tissue compartment. To mitigate inter-scanner and inter-patient variability, SUV metrics were normalized to liver uptake by placing a 1-cm-diameter volume of interest within homogeneous liver parenchyma. The resulting liver-normalized standardized uptake ratio (SUR) metrics were reported as SUR (Mean) and SUR (SD) for each tissue compartment [17, 18]. To mitigate inter-center batch effects arising from differences in PET scanners and acquisition protocols, ComBat harmonization was applied to all PET-derived features using the pyComBat package [19].
We performed supplemental measurement-robustness analyses. These included absolute SUVmean and liver-normalized SUR (Mean) summaries for SAT, IMAT, TAT, and skeletal muscle; depot-resolved metabolic estimates; one-voxel TAT mask erosion; exclusion of high-uptake adipose voxels with SUV > = 2.5 within the CT-defined TAT compartment; and correlations with available clinical/body-habitus variables (Supplemental Tables 1–5).
Double machine learning framework
To evaluate adjusted survival effects of body composition features, we used a double machine learning (DML) framework to quantify adjusted associations for continuous imaging exposures. Each body-composition feature was modeled as a continuous exposure variable, and the outcome was the RMST.
RMST was estimated using jackknife pseudo-observations up to pre-specified truncation horizons (60 months for OS and 36 months for PFS). The resulting pseudo-observations represented individual-level RMST values and were used as the outcome variable in the DML models. For the primary DML analysis, covariates included measured clinical variables and the remaining body-composition features except the target exposure variable. This strategy yielded mutually adjusted conditional associations among correlated imaging phenotypes. Because some body-composition features may share upstream determinants or lie on related biological pathways, these estimates were not interpreted as total effects of isolated tissue phenotypes. The resulting coefficient was interpreted as the DML-adjusted RMST difference associated with a 1-SD higher value of the target exposure, conditional on the measured covariates.
Interpretation of these DML-adjusted estimates requires several assumptions: adequate adjustment for measured confounders, overlap/positivity across the exposure distribution, correct specification or sufficient flexibility of the nuisance models, and absence of major unmeasured confounding. Because this was a retrospective observational study, these assumptions cannot be fully verified; therefore, the estimates are interpreted as adjusted associations rather than definitive causal effects (Supplemental Table 6).
The DML framework consisted of two nuisance models and a final association-estimation step. First, an exposure model was fitted to estimate the conditional expectation of the exposure given covariates, corresponding to the generalized propensity score (GPS) for continuous exposures. Second, an outcome model estimated the conditional expectation of RMST given the same covariates. Both models were implemented using machine learning regressors.
Adjusted association estimates were then obtained using orthogonalized residuals derived from these models, where the residualized outcome was regressed on the residualized exposure. The resulting coefficient represents the DML-adjusted RMST difference, interpreted as the change in RMST (months) associated with a one-standard-deviation increase in the tissue feature.
To further characterize inter-individual heterogeneity, conditional DML-adjusted RMST differences were estimated across patients. Additional methodological details regarding model training, cross-fitting, weighting procedures, and hyperparameter settings are provided in Supplemental Tables 7 and Supplemental Text 1. DAG-informed adjustment-set sensitivity analyses and E-value analyses for unmeasured confounding are described in Supplemental Text 2, Supplemental Fig. 1, and Supplemental Tables 8 and 11 [20].
Because this is a retrospective observational study, the DML-derived estimates should be interpreted as adjusted associations under the above assumptions and not as evidence that changing a body-composition phenotype would causally change survival.
Model construction, validation and incremental-value analysis
To evaluate incremental prognostic value, we performed a model analysis using ridge-penalized Cox regression. Three nested models were evaluated: a clinical model including age, sex, clinical stage, histology, tumor SUVmax, and a tumor-SUVmax missingness indicator; a Clinical + CT model adding CT-derived body-composition features; and a Clinical + CT+PET model further adding PET-derived liver-normalized SUR features. Before modeling, CT-derived body-composition features and PET-derived SUR features were clustered separately according to absolute Spearman correlation, and one representative feature from each cluster was retained. Ridge penalties were selected within inner cross-validation. Internal performance was estimated using 5 repeats of 5-fold nested cross-validation in Center 1, and final models were fitted once in the full Center 1 cohort and evaluated once in the locked Center 2 cohort. For clinical risk visualization, Kaplan-Meier curves were generated using the predicted risk scores obtained from the final model. Time-dependent ROC curves were plotted at prespecified horizons. Additional calibration, decision-curve, Delta C-index, NRI, and IDI analyses were also reported.
Development of local software
To support transparent and reproducible use of the image-processing workflow, we developed a local Windows-compatible research software framework for PET/CT body-composition analysis. The locally run software accepts anonymized CT and PET NIfTI files, performs tissue segmentation, extracts CT-derived volumetric and attenuation features and PET-derived liver-normalized SUR features, displays variables by tissue compartment, and allows CSV export. The framework is also designed to support a research-only risk-stratification module based on the ridge-penalized Cox model, displaying input variables, model-derived risk score and risk stratification results.
Statistical analysis
Statistical analyses were performed using Python 3.10.20. The main Python libraries used were scikit-learn 1.7.2, lifelines 0.30.0, scikit-survival 0.25.0, SimpleITK 2.5.4, neuroCombat 0.2.12, and TotalSegmentator 2.13.0. Cox-model-based E-value sensitivity analyses were performed in R 4.5.3 using the survival package 3.8-6. Continuous variables were summarized as mean ± SD or median (interquartile range) and compared using t tests or Mann-Whitney U tests, as appropriate; categorical variables were compared using chi-square tests. Feature correlations were assessed with Pearson coefficients. For model validation, Harrell’s C-index and time-dependent ROC curves were reported for discrimination. Calibration, Delta C-index, continuous NRI, IDI, and decision curve analyses were summarized in the Supplementary Materials.
Results
Patient characteristics
The patient selection flowchart is presented in (Fig. 1B). A total of 769 patients were incorporated into this study, comprising 555 patients from Center 1 (275 males and 280 females; mean age 61.22 ± 9.97 years) and 214 patients from Center 2 (138 males and 76 females; mean age 64.07 ± 9.53 years). The median interval between PET/CT and surgery were 10.08 and 14.12 days in two centers. In Center 1, 18.91% (105/555) of patients experienced progression or death as PFS events, and 11.17% (62/555) died. The mean OS and PFS time were 49.29 and 45.74 months, respectively. In Center 2, 17.3% (37/214) of patients exhibited disease progression, and 9.8% (21/214) died. The mean OS and PFS time were 39.12 and 36.95 months, respectively.
The data from Center 1 were analyzed using the DML framework and were entirely utilized as the training cohort. Patients from Center 2 were employed as the external validation cohort. There were no significant differences in pathology type, lesion size, lesion location, clinical stage and surgery type among two centers (p > 0.05). Table 1 presents a summary of the clinical characteristics of two centers.
Table 1.
Baseline characteristics of patients with resectable NSCLC in Center 1 and Center 2
| Variable | Overall N = 769 |
Center 1 N = 555 |
Center 2 N = 214 |
statistics | p-value |
|---|---|---|---|---|---|
| Gender | 12.43 | < 0.001 | |||
| female | 357 (46.42%) | 280 (50.45%) | 77 (35.98%) | ||
| male | 412 (53.58%) | 275 (49.55%) | 137 (64.02%) | ||
| Age | 63.00 [56.00, 70.00] | 62.00 [55.00, 69.00] | 65.00 [58.00, 71.00] | 50,675.00 | 0.002 |
| Pathological Type | 3.51 | 0.172 | |||
| Adenocarcinoma | 619 (80.49%) | 453 (81.62%) | 166 (77.57%) | ||
| Squamous carcinoma | 115 (14.95%) | 75 (13.51%) | 40 (18.69%) | ||
| Other | 35 (4.55%) | 27 (4.86%) | 8 (3.74%) | ||
| Lesion Density | 62.16 | < 0.001 | |||
| Solid | 518 (67.36%) | 328 (59.10%) | 190 (88.79%) | ||
| Part-Solid | 174 (22.63%) | 159 (28.65%) | 15 (7.01%) | ||
| Ground Glass Nodule | 77 (10.01%) | 68 (12.25%) | 9 (4.21%) | ||
| Lesion Diameter (cm) | 2.40 [1.80, 3.20] | 2.40 [1.70, 3.30] | 2.40 [2.20, 2.90] | 57,083.50 | 0.404 |
| Clinical Stage | 3.33 | 0.189 | |||
| I | 580 (75.42%) | 409 (73.69%) | 171 (79.91%) | ||
| II | 110 (14.30%) | 86 (15.50%) | 24 (11.21%) | ||
| III | 79 (10.27%) | 60 (10.81%) | 19 (8.88%) | ||
| Location of Lesion | 0.38 | 0.538 | |||
| Left | 326 (42.39%) | 231 (41.62%) | 95 (44.39%) | ||
| Right | 443 (57.61%) | 324 (58.38%) | 119 (55.61%) | ||
| Type of Surgery | 0.87 | 0.646 | |||
| Wedge Resection | 85 (11.05%) | 59 (10.63%) | 26 (12.15%) | ||
| Segmentectomy | 117 (15.21%) | 88 (15.86%) | 29 (13.55%) | ||
| Lobectomy | 567 (73.73%) | 408 (73.51%) | 159 (74.30%) | ||
| 1n (%); Median [IQR] | |||||
2Pearson’s Chi-squared test; Wilcoxon rank sum test
Baseline gender differences in whole body composition
Automated whole-body segmentation revealed pronounced sex-specific differences across most CT- and PET-derived body composition features (Fig. 3A). Male patients demonstrated significantly greater SM volume indices, higher SM density, and increased metabolic activity within SM and IMAT compartments. In contrast, female patients exhibited relatively higher SAT density heterogeneity and distinct adipose distribution patterns.
Fig. 3.

Sex-specific differences in whole-body composition features. (A) Boxplots demonstrate distributions across tissue compartments with statistical significance indicated. Correlation heatmaps of CT-derived morphologic features and PET-derived metabolic features across whole-body tissue compartments. (B, C)
Correlation analysis further demonstrated that CT-based density and volume metrics were strongly interrelated within tissue compartments but showed weaker correlations across compartments (Fig. 3B, C). PET-derived SUR metrics exhibited moderate-to-high intra-compartment correlations (r = 0.4–0.9), indicating consistent metabolic patterns within individual tissue compartments.
DML-adjusted RMST differences of body composition features
Using RMST-GPS-DML, we identified directionally consistent adjusted survival effect estimates for several body composition phenotypes. (Fig. 4) shows the DML-adjusted RMST differences of each feature across OS and PFS. Intermuscular adipose tissue (IMAT) burden emerged as the most prominent adverse body composition feature in the adjusted analysis. Each one-standard-deviation increase in IMAT volume index was associated with a reduction of 4.29 months in OS and 2.74 months in PFS, conditional on measured clinical covariates and other tissue features. Similarly, the IMAT/SM ratio exerted a substantial negative effect, highlighting the importance of ectopic fat infiltration relative to muscle mass rather than absolute fat quantity alone.
Fig. 4.

DML-adjusted RMST differences of PET/CT-derived body-composition features for overall survival and progression-free survival. Forest plots show adjusted restricted mean survival time (RMST) differences, expressed as months per one-standard-deviation increase in each feature, estimated using the restricted mean survival time-generalized propensity score-double machine learning (RMST-GPS-DML) framework. Positive values indicate longer survival, whereas negative values indicate shorter survival. Error bars indicate 95% confidence intervals. Abbreviations: PET/CT, positron emission tomography/computed tomography; DML, double machine learning; RMST, restricted mean survival time; GPS, generalized propensity score; OS, overall survival; PFS, progression-free survival; SM, skeletal muscle; SAT, subcutaneous adipose tissue; IMAT, intermuscular adipose tissue; TAT, torso adipose tissue; SUR, standardized uptake ratio; SD, standard deviation; CI, confidence interval
PET-derived metabolic features showed compartment-specific adjusted associations. TAT SUR (Mean) showed a favorable adjusted association with OS (DML-adjusted RMST difference, + 4.42 months) and PFS (+ 2.72 months). In contrast, SAT SUR (Mean) showed near-null estimates for OS (-0.08 months) and PFS (-0.18 months), and IMAT SUR (Mean) showed weaker or inconsistent estimates (OS, + 0.16 months; PFS, -0.55 months). Depot-resolved metabolic estimates are provided in Supplemental Table 2. Skeletal muscle volume alone showed comparatively modest and often non-significant adjusted survival effects. Indicators of muscle quality, particularly metabolic heterogeneity, demonstrated more consistent associations, reinforcing the concept that muscle function and composition outweigh sheer muscle quantity in prognostic relevance.
Absolute SUVmean and liver-normalized SUR (Mean) distributions for SAT, IMAT, TAT, and skeletal muscle are provided in Supplemental Table 1. High-uptake adipose voxel fractions are reported in Supplemental Table 3. In companion Cox models, the original TAT SUR (Mean) association with OS was HR 0.524 (95% CI, 0.309–0.889; n = 552, 62 events). The association was attenuated after one-voxel ROI erosion (HR 0.748; 95% CI, 0.502–1.115; n = 549, 62 events), but persisted after excluding high-uptake adipose voxels with SUV > = 2.5 (HR 0.501; 95% CI, 0.296–0.847; n = 549, 62 events). Full sensitivity models and clinical correlations are reported in Supplemental Tables 4 and 5.
Sex-stratified DML analyses revealed effect heterogeneity (Tables 2 and 3). In male patients, higher TAT, SAT and IMAT volume indices were associated with shorter OS (DML-adjusted RMST differences, -10.99, -6.20 and − 7.07 months, respectively). In female patients, general adipose volume showed weaker adverse associations, although IMAT volume remained detrimental (-5.03 months). The favorable association of TAT SUR (Mean) with OS was stronger in females than in males (+ 7.77 vs. + 4.89 months), suggesting sex-specific differences in adipose volume and metabolic phenotypes.
Table 2.
Sex-stratified DML-adjusted RMST differences of body composition features on restricted mean overall survival
| Feature | PFS | |||||
|---|---|---|---|---|---|---|
| Male | Female | |||||
| Adjusted RMST difference | CI_low | CI_high | Adjusted RMST difference | CI_low | CI_high | |
| SM Volume Index | -0.222 | -1 | 0.618 | 2.783 | 1.683 | 3.691 |
| SM Density (Mean) | -0.327 | -0.633 | -0.13 | -1.145 | -1.665 | -0.633 |
| SM Density (SD) | 0.243 | 0.021 | 0.478 | 0.838 | 0.012 | 1.333 |
| SAT Volume Index | -6.199 | -7.811 | -4.216 | 0.218 | -0.239 | 0.645 |
| SAT Density (Mean) | -0.882 | -1.471 | -0.571 | -3.142 | -3.959 | -2.37 |
| SAT Density (SD) | 0.413 | 0.186 | 0.689 | 0.36 | 0.004 | 0.649 |
| IMAT Volume Index | -7.073 | -10.756 | -2.828 | -5.032 | -6.153 | -3.393 |
| IMAT/SM Ratio | -6.351 | -8.912 | -3.727 | -0.295 | -1.558 | 0.65 |
| IMAT Density (Mean) | -0.077 | -0.275 | 0.215 | -0.492 | -1.19 | 0.582 |
| IMAT Density (SD) | 0.429 | 0.18 | 0.833 | -1.141 | -1.453 | -0.729 |
| TAT Volume Index | -10.99 | -14.61 | -6.803 | -0.097 | -2.287 | 1.964 |
| TAT/SAT Ratio | -7.119 | -9.405 | -3.569 | 0.277 | -0.24 | 0.736 |
| TAT Density (Mean) | -2.397 | -4.264 | -0.277 | -1.98 | -3.062 | -0.922 |
| TAT Density (SD) | 0.509 | 0.223 | 0.86 | 1.871 | 0.464 | 2.789 |
| Bone Density (Mean) | -2.822 | -3.192 | -2.025 | -2.544 | -3.104 | -1.798 |
| Bone Density (SD) | 2.39 | 1.52 | 2.998 | 1.004 | 0.724 | 1.305 |
| SM SUR (Mean) | 4.333 | 2.655 | 5.733 | -1.686 | -2.958 | -0.408 |
| SM SUR (SD) | 1.915 | 0.773 | 2.761 | -1.224 | -2.176 | -0.455 |
| SAT SUR (Mean) | 1.987 | 0.817 | 3.053 | -4.637 | -5.458 | -3.649 |
| SAT SUR (SD) | 0.317 | -0.563 | 1.372 | 3.411 | 2.494 | 4.435 |
| IMAT SUR (Mean) | 1.273 | 0.069 | 2.255 | -1.479 | -1.974 | -0.779 |
| IMAT SUR (SD) | 4.205 | 2.189 | 5.823 | 1.125 | 0.361 | 1.766 |
| TAT SUR (Mean) | 4.885 | 3.23 | 6.949 | 7.774 | 6.185 | 9.308 |
| TAT SUR (SD) | -2.268 | -3.02 | -1.08 | -2.256 | -3.47 | -1.33 |
| Abbreviations SUR standardized uptake ratio, SM skeletal muscle, SAT subcutaneous adipose tissue, IMAT intermuscular adipose tissue, TAT torso adipose tissue, OS overall survival, PFS progression-free survival | ||||||
Table 3.
Sex-stratified DML-adjusted RMST differences of body composition features on restricted mean progression-free survival
| Feature | PFS | |||||
|---|---|---|---|---|---|---|
| Male | Female | |||||
| Adjusted RMST difference | CI_low | CI_high | Adjusted RMST difference | CI_low | CI_high | |
| SM Volume Index | 0.13 | -0.519 | 0.795 | 2.429 | 1.74 | 2.947 |
| SM Density (Mean) | -0.564 | -0.881 | -0.458 | -1.464 | -2.031 | -0.784 |
| SM Density (SD) | 0.438 | 0.342 | 0.618 | 0.524 | -0.047 | 0.817 |
| SAT Volume Index | -4.803 | -5.795 | -3.972 | -0.948 | -1.401 | -0.238 |
| SAT Density (Mean) | -0.789 | -1.541 | -0.636 | -1.952 | -2.586 | -1.466 |
| SAT Density (SD) | 0.499 | 0.368 | 0.67 | -0.591 | -1.059 | -0.037 |
| IMAT Volume Index | -4.338 | -5.958 | -2.811 | -3.211 | -4.033 | -2.303 |
| IMAT/SM Ratio | -3.657 | -4.461 | -2.643 | -2.863 | -4.236 | -0.968 |
| IMAT Density (Mean) | -0.276 | -0.438 | -0.131 | -0.066 | -0.49 | 0.528 |
| IMAT Density (SD) | 0.464 | 0.335 | 0.673 | -0.128 | -0.317 | 0.097 |
| TAT Volume Index | -2.093 | -4.069 | 0.533 | -1.813 | -3.357 | -0.746 |
| TAT/SAT Ratio | -3.935 | -5.198 | -2.55 | 0.474 | 0.066 | 0.971 |
| TAT Density (Mean) | -0.156 | -1.292 | 1.314 | -2.776 | -3.706 | -1.871 |
| TAT Density (SD) | 0.604 | 0.487 | 0.873 | 0.419 | -0.426 | 1.198 |
| Bone Density (Mean) | -2.807 | -3.179 | -2.049 | -2.017 | -2.348 | -1.542 |
| Bone Density (SD) | 1.069 | 0.562 | 1.611 | 0.071 | -0.127 | 0.269 |
| SM SUR (Mean) | 1.426 | 0.679 | 2.233 | -1.838 | -2.59 | -1.091 |
| SM SUR (SD) | -0.326 | -1.148 | 0.37 | 0 | -0.666 | 0.669 |
| SAT SUR (Mean) | 0.37 | -0.262 | 1.042 | -2.483 | -3.01 | -1.826 |
| SAT SUR (SD) | -0.763 | -1.298 | 0.029 | 0.709 | 0.206 | 1.448 |
| IMAT SUR (Mean) | -0.429 | -1.05 | 0.005 | -1.52 | -1.806 | -1.159 |
| IMAT SUR (SD) | 1.537 | 0.563 | 2.377 | 0.906 | 0.455 | 1.334 |
| TAT SUR (Mean) | 0.436 | -0.64 | 2.106 | 5.077 | 4.14 | 5.965 |
| TAT SUR (SD) | -1.525 | -2.03 | -0.743 | -1.052 | -1.838 | -0.432 |
Abbreviations SUR standardized uptake ratio, SM skeletal muscle, SAT subcutaneous adipose tissue, IMAT intermuscular adipose tissue, TAT torso adipose tissue, OS overall survival, PFS progression-free survival
In DAG-informed adjustment-set sensitivity analyses, the TAT SUR (Mean) association remained directionally favorable across alternative adjustment sets (Supplemental Table 8). The association was weak in clinical-only models but became stronger after accounting for body-composition context.
Heterogeneity in DML-adjusted RMST associations of body composition features on survival
Conditional adjusted RMST difference distributions (Fig. 5; Supplemental Tables 9 and 10) showed substantial inter-individual heterogeneity in the associations between body-composition features and survival. Compared with the population-level DML-adjusted RMST differences, the conditional distributions displayed wide bilateral tails, indicating that the magnitude and direction of these adjusted associations varied across patients.
Fig. 5.

Distributions of conditional DML-adjusted RMST differences for OS and PFS, demonstrating inter-individual heterogeneity in conditional adjusted survival associations to body composition features
For IMAT burden, the conditional adjusted RMST difference distribution was predominantly shifted toward negative values for both OS and PFS, with a dense left tail. This pattern indicates that higher IMAT burden was generally associated with shorter postoperative survival, although a small subset of patients showed neutral or mildly positive conditional adjusted RMST differences.
Conversely, adipose metabolic features, including TAT SUR(mean) and SAT SUR(SD), showed right-shifted conditional adjusted RMST difference distributions. This pattern suggests that the favorable adjusted association between adipose metabolic activity and survival was more pronounced in a subset of patients, while remaining descriptive rather than implying individualized causal benefit.
In the E-value sensitivity analysis, companion Cox models were fitted for the headline DML associations. For OS, 1-SD higher IMAT volume index had a companion Cox HR of 1.955 (95% CI, 0.961–3.981), corresponding to an E-value of 3.322 for the point estimate. For OS TAT SUR (Mean), the companion Cox HR was 0.528 (95% CI, 0.311–0.897), corresponding to an E-value of 3.197 for the point estimate and 1.474 for the confidence-limit closest to the null. Full E-value results are provided in Supplemental Table 11.
Model validation and incremental-value analysis
The Center 1 development cohort included 555 patients with 62 deaths and 105 progression/death events. Correlation clustering reduced CT-derived body-composition features to 10 representatives and PET-derived SUR features to 4 representatives before ridge-Cox modeling. The final Clinical + CT+PET model was then used for clinical risk visualization (Fig. 6). High-risk groups showed shorter OS and PFS in the development cohort (OS log-rank P = 0.005; PFS P < 0.01) and in the external validation cohort (OS log-rank P = 0.046; PFS P < 0.01).
Fig. 6.

Clinical risk stratification using the final reduced ridge-penalized Cox model. Kaplan-Meier curves for OS and PFS in the development and external validation cohorts (A-D). Time-dependent ROC curves for OS and PFS in the development and external validation cohorts are shown at prespecified horizons (E-H)
Time-dependent ROC curves showed moderate discrimination. In the external cohort, the final Clinical + CT+PET model achieved AUCs of 0.761, 0.644, and 0.627 for 1-, 3-, and 5-year OS, and 0.725, 0.652, and 0.666 for 1-, 2-, and 3-year PFS, respectively. Harrell C-indices for the final model were 0.703 in development and 0.653 in external validation for OS, and 0.678 in development and 0.656 in external validation for PFS. In the formal nested comparison, incremental value over the clinical model was modest and endpoint-dependent; external Delta C-index, continuous NRI, IDI, calibration, and decision-curve results are summarized in Supplemental Tables 12 and Supplemental Fig. 2.
A local body-composition analysis software framework was prepared to demonstrate reproducible deployment of the segmentation and feature-extraction workflow (https://github.com/AI-Lung/AI-Lung-PETCT-Lung-Body-Composition-Survival; Fig. 7). The software displays extracted CT- and PET-derived body-composition variables by tissue compartment, and provides a framework for local research-only risk-score calculation using the model.
Fig. 7.

Local PET/CT body-composition analysis software. The interface accepts anonymized CT and PET NIfTI files, performs TotalSegmentator-based tissue segmentation, displays extracted CT- and PET-derived body-composition parameters by compartment, exports the resulting feature table, and provides a framework for local research-only risk-score calculation
Discussion
This study suggests that body composition in resectable NSCLC may be better characterized as a biologically heterogeneous imaging phenotype rather than as a set of static single-slice morphologic measurements. IMAT burden emerged as a consistent adverse factor for both OS and PFS, while TAT SUR (Mean) showed an exploratory favorable adjusted association with OS. Pronounced sex-specific differences were also observed: male patients appeared more vulnerable to adipose volume accumulation, whereas female patients showed stronger favorable adjusted associations with adipose metabolic activity. These findings suggest that PET/CT-derived body-composition phenotypes capture clinically relevant host-tissue information, although biological interpretation remains exploratory.
IMAT as a central driver of adverse prognosis
Rather than establishing IMAT as a newly recognized biomarker, our results extend prior observations by showing that whole-body PET/CT can jointly quantify ectopic adiposity, tissue attenuation, and metabolic activity, and that these features can be incorporated into an adjusted survival modeling framework. One of the most consistent findings in our cohort is the strong adverse association between IMAT burden and both OS and PFS. This observation is further supported by the finding that increased TAT volume was also associated with shorter survival. Unlike SAT, IMAT reflects ectopic lipid accumulation within skeletal muscle and is closely linked to insulin resistance, chronic inflammation, and impaired muscle contractility [21, 22]. Our findings are consistent with prior work linking IMAT to mortality in NSCLC; however, those analyses relied on two-dimensional CT and conventional Cox models. In contrast, we used three-dimensional whole-body analysis with double machine learning, enabling more comprehensive and robust estimation [23]. Even after adjustment for muscle mass and adipose indices, higher IMAT was correlated with worse OS. Sex-stratified analyses further demonstrated that the adverse effects of adiposity were largely driven by male patients. In men, increases across multiple fat depots - including SAT, TAT, and IMAT - were associated with substantial reductions in survival time, with DML-adjusted RMST reductions ranging from approximately 6 to 11 months.
Mechanistically, IMAT is metabolically active and can secrete pro-inflammatory cytokines such as IL-6 and TNF-α, which may promote tumor progression and impair antitumor immunity [24]. Although the exact cytokine milieu of IMAT in NSCLC remains incompletely defined, ectopic adipose deposition is increasingly recognized as a contributor to systemic metabolic dysregulation in cancer [25].
Muscle quantity versus muscle quality
Unlike conventional sarcopenia theories, skeletal muscle volume exhibited generally minor and frequently insignificant effects when ectopic fat infiltration and metabolic characteristics were included. The finding aligns with growing radiological evidence indicating that muscle quality, rather than quantity, is more significantly associated with oncologic outcomes [26–28]. In this study, the IMAT/SM ratio serves as a volumetric indicator of myosteatosis and may represent a macroscopic manifestation of deteriorating muscle quality. Our results suggest that the adverse survival associations historically attributed to “sarcopenia” may partly reflect the presence of ectopic fat infiltration rather than simple reductions in contractile muscle mass. Consequently, greater muscle volume does not necessarily confer a survival advantage if accompanied by substantial IMAT burden. The wide distribution of conditional adjusted RMST differences further indicates substantial heterogeneity, implying that the favorable association of greater muscle mass with survival depends on the underlying metabolic context, including fat infiltration and systemic inflammatory status.
TAT metabolic activity and measurement robustness
An additional finding was the favorable adjusted association between TAT SUR (Mean) and OS. At a broad level, this observation is relevant to the literature on the obesity paradox in cancer, including lung cancer, in which higher BMI has sometimes been associated with improved outcomes [29, 30]. BMI, however, is a crude anthropometric measure and cannot capture tissue-specific metabolic activity or distinguish between adipose compartments. PET-derived body-composition metrics may therefore provide functional information that complements CT-based anatomical quantification [31, 32].
In this study, TAT was defined as total torso fat within the thoracic, abdominal, and pelvic cavities, excluding subcutaneous fat, based on the TotalSegmentator torso_fat label and physician review. It should not be equated with conventional VAT. Prior VAT-focused PET studies have reported adverse prognostic associations of higher VAT FDG uptake, including poorer recurrence-free survival in colorectal cancer and breast cancer [33, 34]. Because those studies used 2D segmentation and VAT-specific masks, their findings are not directly comparable with the TAT phenotype analyzed here. In our depot-resolved analyses, SAT SUR (Mean) showed near-null associations, suggesting that the TAT signal was not simply driven by subcutaneous fat uptake. The association also persisted after excluding high-uptake adipose voxels with SUV > = 2.5, arguing against dominance by a small fraction of high-uptake voxels. However, it was attenuated after one-voxel ROI erosion, indicating sensitivity to boundary definition and possible partial-volume effects. We therefore interpret TAT SUR (Mean) as an exploratory imaging phenotype rather than evidence of protective adipose biology.
Measurement-robustness analyses supported this cautious interpretation. The favorable TAT SUR (Mean) association was not driven solely by high-uptake adipose voxels, but its attenuation after ROI erosion indicates that compartment-boundary definition may influence the estimate. These results strengthen the need for external validation with dedicated adipose-depot segmentation and more complete metabolic, nutritional, and inflammatory covariates.
Adjustment-set sensitivity analyses provided a similar message. The TAT SUR (Mean) association was not abolished after excluding adipose volume/ratio terms, but it was weak in clinical-only models and became more apparent after accounting for body-composition context. This pattern supports cautious interpretation as a conditional imaging phenotype rather than as a total or causal protective effect.
Sex-specific biological modulation
Sex-stratified analyses revealed substantial effect modification by sex. Male patients appeared particularly vulnerable to the adverse effects of adipose accumulation across multiple depots - including SAT, TAT, and IMAT - whereas female patients showed stronger favorable adjusted associations for adipose metabolic metrics. These differences may reflect sex-related variation in fat distribution, hormonal regulation, and immune responses. Estrogen has been shown to promote subcutaneous fat deposition and enhance lipid oxidation, potentially mitigating the metabolic consequences of ectopic fat accumulation [35, 36]. In addition, sex-based differences in inflammatory signaling and immune checkpoint regulation may influence the interaction between metabolic tissues and tumor biology [37].
Disease specificity and generalizability
The extent to which these associations are specific to NSCLC remains uncertain. Some body-composition phenotypes, particularly IMAT accumulation and impaired muscle quality, likely reflect systemic processes such as ectopic lipid deposition, insulin resistance, chronic inflammation, frailty, and reduced metabolic reserve. These processes are not unique to NSCLC and may also be relevant in other cancers, nonmalignant chronic diseases, and aging populations. Therefore, the adverse association between IMAT burden and survival may represent a broader host-related risk phenotype rather than an NSCLC-specific mechanism alone. However, the magnitude and direction of these associations may vary substantially by disease context. In resectable NSCLC, prognosis is influenced by smoking-related inflammation, pulmonary reserve, tumor stage, surgical stress, perioperative complications, and cancer-associated cachexia. PET-derived metabolic features may be even more context-dependent, because FDG uptake is affected by tumor biology, systemic inflammation, glucose metabolism, image acquisition protocols, and treatment setting. Accordingly, our findings should be considered directly applicable only to patients with resectable NSCLC undergoing preoperative PET/CT and curative-intent surgery. Validation in other malignancies, benign diseases, and general-population cohorts is required before broader generalization.
Clinical and methodological implications
The model analysis provides a more clinically grounded picture of potential translation. The final model supported visual risk stratification in both development and external cohorts, and the external C-indices were in a moderate range. However, nested comparisons showed that incremental value over clinical variables was endpoint-dependent: body-composition features provided a modest numerical improvement for OS but not for PFS. These findings support local software-based feature extraction and research-oriented risk stratification, while indicating that prospective validation is required before using the model for treatment decision-making.
Therefore, the present results should be viewed primarily as evidence that PET/CT-derived body-composition phenotypes can generate interpretable prognostic signals and clinically testable hypotheses about host-tumor metabolic status. Larger prospective cohorts with standardized acquisition, complete tumor metabolic variables, and predefined clinical decision thresholds are needed to determine whether PET/CT body-composition phenotyping improves risk stratification beyond established clinicopathologic models.
Limitations and future directions
Our study has several limitations. First, its retrospective observational design may introduce selection bias and residual confounding. Although DML flexibly adjusted for measured covariates and E-values provided quantitative context, these analyses do not establish causality. Unmeasured factors such as smoking intensity, comorbidity burden, weight trajectory, nutritional status, performance status, serum glucose, CRP, and NLR may remain. Second, the tissue_4_types task does not provide dedicated VAT or BAT masks. TAT therefore should not be interpreted as conventional VAT, and VAT- versus SAT-specific metabolic effects could not be directly assessed. Third, correlated body-composition phenotypes may involve confounding, mediation, or collider structures; fully adjusted DML estimates should therefore be interpreted as conditional adjusted associations rather than total effects. Fourth, the sample size, particularly for subgroup analyses and model validation, may limit stability and precision. Fifth, because this cohort was limited to patients with resectable NSCLC, the observed associations should not be extrapolated to other cancers, nonmalignant diseases, or the general population without disease-specific validation. Finally, residual scanner/protocol variability may persist despite SUR normalization and ComBat harmonization, and fixed RMST horizons may not capture longer-term outcome differences. Prospective multicenter studies are needed to validate these findings, assess their generalizability across disease contexts, and determine their incremental value beyond established clinical models.
Conclusions
This study suggests that ectopic IMAT accumulation is associated with shorter postoperative survival in resectable NSCLC, whereas TAT SUR (Mean) showed an exploratory favorable adjusted association with OS. By integrating whole-body PET/CT phenotyping with adjusted survival modeling, our approach provides interpretable prognostic information and a reproducible software framework for local feature extraction and research-oriented risk stratification.
Supplementary Information
Below is the link to the electronic supplementary material.
Author contribution
Weihao Zhai contributed to study conception, study design, data analysis and drafting of the manuscript. Ruoyao Wang, Mengmeng Ye, Xiaolin Li, Yi Wang, Taohu Zhou, Xiuxiu Zhou and Qianxi Jin contributed to data collection, image processing and validation. Ziwei Zhang and Li Fan contributed to study supervision, interpretation of results and critical revision of the manuscript. All authors read and approved the final manuscript.
Funding
This work was supported by the National Natural Science Foundation of China (No. 82430065), the Excellent Health Sector Program of Shanghai Municipal Health Commission (20254Z0003), and the Shanghai “Rising Stars of Medical Talent” Youth Development Program for Outstanding Youth Medical Talents (SHWSRS 2025-71).
Data availability
The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request, subject to institutional and ethical restrictions.
Declarations
Ethics approval
This retrospective study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Research Ethics Committee of the Institutional Review Boards of the two participating centers (approval number 2024SL044), and this approval covered both centers.
Consent to participate
The Research Ethics Committee of the Institutional Review Boards of the two participating centers waived the requirement for informed consent because of the retrospective study design.
Competing interests
The authors have no relevant financial or non-financial interests to disclose.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A, Cancer statistics. 2025. CA Cancer J Clin. 2025;75(1):10–45. 10.3322/caac.21871 [DOI] [PMC free article] [PubMed]
- 2.Goldstraw P, Chansky K, Crowley J, et al. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol. 2016;11(1):39–51. 10.1016/j.jtho.2015.09.009. [DOI] [PubMed] [Google Scholar]
- 3.Detterbeck FC, Boffa DJ, Kim AW, Tanoue LT. The Eighth Edition Lung Cancer Stage Classification. Chest. 2017;151(1):193–203. 10.1016/j.chest.2016.10.010. [DOI] [PubMed] [Google Scholar]
- 4.Martin L, Birdsell L, MacDonald N, et al. Cancer Cachexia in the Age of Obesity: Skeletal Muscle Depletion Is a Powerful Prognostic Factor, Independent of Body Mass Index. J Clin Oncol. 2013;31(12):1539–47. 10.1200/JCO.2012.45.2722. [DOI] [PubMed] [Google Scholar]
- 5.Aubrey J, Esfandiari N, Baracos VE, et al. Measurement of skeletal muscle radiation attenuation and basis of its biological variation. Acta Physiol. 2014;210(3):489–97. 10.1111/apha.12224. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Prado CM, Lieffers JR, McCargar LJ, et al. Prevalence and clinical implications of sarcopenic obesity in patients with solid tumours of the respiratory and gastrointestinal tracts: a population-based study. Lancet Oncol. 2008;9(7):629–35. 10.1016/S1470-2045(08)70153-0. [DOI] [PubMed] [Google Scholar]
- 7.Mourtzakis M, Prado CMM, Lieffers JR, Reiman T, McCargar LJ, Baracos VE. A practical and precise approach to quantification of body composition in cancer patients using computed tomography images acquired during routine care. Appl Physiol Nutr Metab. 2008;33(5):997–1006. 10.1139/H08-075. [DOI] [PubMed] [Google Scholar]
- 8.MacDonald AJ, Greig CA, Baracos V. The advantages and limitations of cross-sectional body composition analysis. Curr Opin Support Palliat Care. 2011;5(4):342–9. 10.1097/SPC.0b013e32834c49eb. [DOI] [PubMed] [Google Scholar]
- 9.Wasserthal J, Breit HC, Meyer MT, et al. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell. 2023;5(5):e230024. 10.1148/ryai.230024. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Bao J, Xiong M, Zheng M, Huang P, Lin X. Tumour and adipose tissue uptake on 18F-Fluorodeoxyglucose (18F-FDG) positron emission tomography/computed tomography (PET/CT) predict immunotherapy response in recurrent or metastatic nasopharyngeal carcinoma. Clin Radiol. 2025;87:106959. 10.1016/j.crad.2025.106959. [DOI] [PubMed] [Google Scholar]
- 11.Fearon K, Strasser F, Anker SD, et al. Definition and classification of cancer cachexia: an international consensus. Lancet Oncol. 2011;12(5):489–95. 10.1016/S1470-2045(10)70218-7. [DOI] [PubMed] [Google Scholar]
- 12.Feuerriegel S, Frauen D, Melnychuk V, et al. Causal machine learning for predicting treatment outcomes. Nat Med. 2024;30(4):958–68. 10.1038/s41591-024-02902-1. [DOI] [PubMed] [Google Scholar]
- 13.Wang X, Liu Y, Qin G, Yu Y. Robust double machine learning model with application to omics data. BMC Bioinformatics. 2024;25(1):355. 10.1186/s12859-024-05975-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Westling T, Luedtke A, Gilbert PB, Carone M. Inference for Treatment-Specific Survival Curves Using Machine Learning. J Am Stat Assoc. 2024;119(546):1541–53. 10.1080/01621459.2023.2205060. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Royston P, Parmar MK. Restricted mean survival time: An alternative to the hazard ratio for the design and analysis of randomized trials with a time-to-event outcome. BMC Med Res Methodol. 2013;13(1):152. 10.1186/1471-2288-13-152. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Borys K, Lodde G, Livingstone E, et al. Fully volumetric body composition analysis for prognostic overall survival stratification in melanoma patients. J Transl Med. 2025;23(1):532. 10.1186/s12967-025-06507-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Boellaard R, Delgado-Bolton R, Oyen WJG, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. Eur J Nucl Med Mol Imaging. 2015;42(2):328–54. 10.1007/s00259-014-2961-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Wahl RL, Jacene H, Kasamon Y, Lodge MA. From RECIST to PERCIST: Evolving Considerations for PET Response Criteria in Solid Tumors. J Nucl Med. 2009;50(Suppl 1):S122–50. 10.2967/jnumed.108.057307. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Orlhac F, Boughdad S, Philippe C, et al. A Postreconstruction Harmonization Method for Multicenter Radiomic Studies in PET. J Nucl Med. 2018;59(8):1321–8. 10.2967/jnumed.117.199935. [DOI] [PubMed] [Google Scholar]
- 20.VanderWeele TJ, Ding P. Sensitivity Analysis in Observational Research: Introducing the E-Value. Ann Intern Med. 2017;167(4):268–74. 10.7326/M16-2607. [DOI] [PubMed] [Google Scholar]
- 21.Deng HY, Hou L, Zha P, Huang KL, Peng L. Sarcopenia is an independent unfavorable prognostic factor of non-small cell lung cancer after surgical resection: A comprehensive systematic review and meta-analysis. Eur J Surg Oncol. 2019;45(5):728–35. 10.1016/j.ejso.2018.09.026. [DOI] [PubMed] [Google Scholar]
- 22.Park EJ, Lee JH, Yu GY, et al. Dietary and Genetic Obesity Promote Liver Inflammation and Tumorigenesis by Enhancing IL-6 and TNF Expression. Cell. 2010;140(2):197–208. 10.1016/j.cell.2009.12.052. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Huang Y, Cun H, Mou Z, et al. Multiparameter body composition analysis on chest CT predicts clinical outcomes in resectable non-small cell lung cancer. Insights Imaging. 2025;16(1):32. 10.1186/s13244-025-01910-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Baracos VE, Martin L, Korc M, Guttridge DC, Fearon KCH. Cancer-associated cachexia. Nat Rev Dis Primer. 2018;4(1):17105. 10.1038/nrdp.2017.105. [DOI] [PubMed] [Google Scholar]
- 25.Sabatino A, Cordeiro AC, Prado CM, Lindholm B, Stenvinkel P, Avesani CM. Myosteatosis is associated with adiposity, metabolic derangements and mortality in patients with chronic kidney disease. Eur J Clin Nutr. 2025;79(5):475–83. 10.1038/s41430-024-01551-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Mantz L, Mercaldo ND, Simon J, et al. Preoperative Chest CT Myosteatosis Indicates Worse Postoperative Survival in Stage 0–IIB Non–Small Cell Lung Cancer. Radiology. 2025;314(2). 10.1148/radiol.240282. [DOI] [PubMed]
- 27.Costa-Pereira JP, Macêdo IF, Barbosa GA, et al. Myosteatosis is associated with advanced TNM stage and tumor features in young females with breast cancer. BMC Cancer. 2025;25(1):1576. 10.1186/s12885-025-15047-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Bradshaw PT. Body composition and cancer survival: a narrative review. Br J Cancer. 2024;130(2):176–83. 10.1038/s41416-023-02470-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Cheng E, Kirley J, Cespedes Feliciano EM, Caan BJ. Adiposity and cancer survival: a systematic review and meta-analysis. Cancer Causes Control. 2022;33(10):1219–46. 10.1007/s10552-022-01613-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Cespedes Feliciano EM, Kroenke CH, Caan BJ. The obesity paradox in Cancer: How Important Is Muscle? Annu Rev Nutr. 2018;38(1):357–79. 10.1146/annurev-nutr-082117-051723. [DOI] [PubMed] [Google Scholar]
- 31.Karaçelik T, Kaya B, Korkmaz M, et al. Prognostic significance of adipose tissue distribution and metabolic activity in PET/CT in Patients with metastatic colorectal cancer. J Gastrointest Cancer. 2023;54(2):456–66. 10.1007/s12029-022-00819-x. [DOI] [PubMed] [Google Scholar]
- 32.Quail DF, Dannenberg AJ. The obese adipose tissue microenvironment in cancer development and progression. Nat Rev Endocrinol. 2019;15(3):139–54. 10.1038/s41574-018-0126-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Yoo ID, Lee SM, Lee JW, Baek MJ, Ahn TS. Usefulness of metabolic activity of adipose tissue in FDG PET/CT of colorectal cancer. Abdom Radiol (NY). 2018;43(8):2052–9. 10.1007/s00261-017-1418-7. [DOI] [PubMed] [Google Scholar]
- 34.Kim HJ, Kim D, Bae SJ, et al. 18F-FDG uptake of visceral adipose tissue on preoperative PET/CT as a predictive marker for breast cancer recurrence. Sci Rep. 2022;12(1):21109. 10.1038/s41598-022-25540-4. Published 2022 Dec 6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Vedire Y, Kalvapudi S, Yendamuri S. Obesity and lung cancer—a narrative review. J Thorac Dis. 2023;15(5):2806–23. 10.21037/jtd-22-1835. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Mosleh B, Sarova P, Zehetmayer S, et al. Sex-based differences in lung cancer susceptibility and molecular genetics in the 2020s. Heliyon. 2025;11(2):e42089. 10.1016/j.heliyon.2025.e42089. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Jiang M, Fares AF, Shepshelovich D, et al. The relationship between body-mass index and overall survival in non-small cell lung cancer by sex, smoking status, and race: A pooled analysis of 20,937 International lung Cancer consortium (ILCCO) patients. Lung Cancer. 2021;152:58–65. 10.1016/j.lungcan.2020.11.029. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request, subject to institutional and ethical restrictions.
