Abstract
Background and objectives
Brain metastasis (BM) significantly affects the prognosis of non-small cell lung cancer (NSCLC) patients. Increasing evidence suggests that adipose tissue influences cancer progression and metastasis. This study aimed to develop a predictive nomogram integrating mediastinal fat area (MFA) and deep learning (DL)-derived tumor characteristics to stratify postoperative BM risk in NSCLC patients.
Materials and methods
A retrospective cohort of 585 surgically resected NSCLC patients was analyzed. Preoperative computed tomography (CT) scans were utilized to quantify MFA using ImageJ software (radiologist-validated measurements). Concurrently, a DL algorithm extracted tumor radiomic features, generating a deep learning brain metastasis score (DLBMS). Multivariate logistic regression identified independent BM predictors, which were incorporated into a nomogram. Model performance was assessed via area under the receiver operating characteristic curve (AUC), calibration plots, integrated discrimination improvement (IDI), net reclassification improvement (NRI), and decision curve analysis (DCA).
Results
Multivariate analysis identified N stage, EGFR mutation status, MFA, and DLBMS as independent predictors of BM. The nomogram achieved superior discriminative capacity (AUC: 0.947 in the test set), significantly outperforming conventional models. MFA contributed substantially to predictive accuracy, with IDI and NRI values confirming its incremental utility (IDI: 0.123, P < 0.001; NRI: 0.386, P = 0.023). Calibration analysis demonstrated strong concordance between predicted and observed BM probabilities, while DCA confirmed clinical net benefit across risk thresholds.
Conclusion
This DL-enhanced nomogram, incorporating MFA and tumor radiomics, represents a robust and clinically useful tool for preoperative prediction of postoperative BM risk in NSCLC. The integration of adipose tissue metrics with advanced imaging analytics advances personalized prognostic assessment in NSCLC patients.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12885-025-14466-5.
Keywords: Brain metastases, Non-small-cell lung cancer, Deep learning, Mediastinal fat area, Nomogram
Introduction
Brain metastasis (BM) occurs in over 10% of non-small cell lung cancer (NSCLC) patients following surgical resection, with untreated cases exhibiting a median survival of merely 1–2 months [1, 2]. Consequently, early preoperative prediction of postoperative BM is critical for optimizing therapeutic decision-making in NSCLC.
While CT-based texture analysis quantifies tumor heterogeneity [3], conventional radiomics approaches often fail to capture complex pathophysiological signatures [4]. Deep learning (DL), in contrast, demonstrates superior capabilities in autonomously extracting high-dimensional features from raw imaging data, enabling robust prediction of tumor behavior with enhanced accuracy and generalizability [5]. Prior studies have validated DL’s utility in forecasting occult lymph node metastasis, overall survival, and treatment response in NSCLC [5–8]; however, its application for preoperative prediction of postoperative BM risk stratification remains unexplored.
Adipose tissue, a key determinant of body composition, actively participates in oncogenesis and metastatic progression [9–11]. Although the L3-level cross-sectional area on abdominal CT serves as the gold standard for body composition assessment [12], this metric is frequently unavailable in lung cancer cohorts due to inconsistent L3 inclusion in routine chest CT protocols [13]. Mediastinal fat, a chest cavity component increasingly recognized for its prognostic relevance in NSCLC [14], reflects systemic metabolic and inflammatory perturbations associated with tumor aggressiveness [15, 16].
This study pioneers the integration of mediastinal fat area (MFA) quantification with DL-derived tumor radiomics to construct a predictive nomogram for preoperative assessment of postoperative BM risk in resected NSCLC. By synergizing adipose biology with advanced imaging analytics, our model aims to refine risk stratification and guide personalized surveillance strategies.
Materials and methods
Study design and participants
The institutional review board of Harbin Medical University Cancer Hospital approved this study (approval number: YD2024–06). Since the study was designed retrospectively, the requirement for informed consent was waived. We included 585 NSCLC patients who underwent surgical resection between January 2015 and November 2018. Participants were randomized into training (n = 390) and test (n = 195) sets using a 2:1 ratio. The model framework is depicted in Fig. 1. Exclusion criteria and workflow are detailed in Supplementary Figure S1.
Fig. 1.
A model framework diagram
Clinical information collection
The data on clinical information and biochemical parameters were obtained through electronic health records, including age, eastern cooperative oncology group performance status (ECOG), sex, smoking history, cytokeratin 19 fragment 21 − 1 (CYFRA21-1) carcinoembryonic antigen (CEA), body mass index (BMI), T stage, N stage, pathological type, epidermal growth factor receptor (EGFR) mutation status, neuron-specific enolase (NSE), and anaplastic lymphoma kinase (ALK) mutation status. The diagnosis of BM required MRI confirmation. Follow-up concluded at either patient death or November 2023.
Mediastinal fat area quantification
Preoperative chest CT scans (≤ 1 month before surgery; parameters in Table S1, Supplementary Methods 1 and 2) were analyzed using ImageJ software. MFA was quantified as a two-dimensional measurement at the first layer upward of the aortic arch, with anatomical boundaries defined as follows: anteriorly by the sternum, posteriorly by vertebral bodies, and laterally by mediastinal pleura and great vessels (HU range: -200 to -40) [17] (Figure S2). This standardized axial plane was selected to ensure consistent cross-sectional assessment across all subjects. Two blinded radiologists performed measurements (intraclass correlation coefficient: 0.92).
Deep learning feature extraction
A 3D convolutional neural network (CNN), pretrained on 50,000 annotated thoracic CT scans, was fine-tuned to extract tumor radiomic features (shape, texture, intensity, heterogeneity) from manually segmented volumes of interest (VOIs). Preprocessing included isotropic resampling (1 × 1 × 1 mm³) and intensity normalization (Supplementary Method 3). Measurements were taken by two blinded radiologists (intraclass correlation coefficient: 0.90).
Model development and optimization
VOIs were processed as three-channel 3D inputs (Supplementary Method 4). We evaluated six different CNN models: DenseNet-121/169, ResNet-50/101, and ResNext-50/101. DenseNet-121 achieved optimal performance (AUC: 0.844) and was selected to generate the DL-based brain metastasis score (DLBMS) (Supplementary Method 5). To address class imbalance, we applied epoch-wise positive case oversampling with a 1:1 ratio. We used gradient-weighted class activation mapping (Grad-CAM) to visualize the regions in CT images that were crucial for predicting patients with different outcomes.
Nomogram construction and validation
Multivariate logistic regression identified independent BM predictors (P < 0.05), which were integrated into a nomogram. The nomogram predicts 5-year postoperative BM risk over the entire follow-up period (median duration: 56 months, range: 3–60 months). Four comparator models were developed: clinical variables only, MFA only, MFA + DLBMS, and Clinical + MFA. Performance was assessed via AUC, calibration curves, integrated discrimination improvement (IDI), net reclassification improvement (NRI), and decision curve analysis (DCA).
Prognostic stratification
Patients were dichotomized into high-/low-risk groups using nomogram-derived Youden index thresholds.
Statistical analysis
Categorical variables were compared via χ² tests, continuous variables via t-tests/Mann-Whitney U tests. Bootstrap resampling (1000 iterations) generated 95% CIs. Survival differences were evaluated using Kaplan-Meier/log-rank methods. Analyses utilized Python 3.9.13 and R 4.0.5 (significance: two-tailed P < 0.05).
Results
Cohort characteristics
Baseline characteristics were balanced across training/test sets (Table 1). BM patients exhibited elevated CEA (P < 0.001), preoperative MFA (P < 0.001), advanced T/N stages (P = 0.003/< 0.001), and higher EGFR mutation prevalence (P = 0.041; Table S2).
Table 1.
Baseline characteristics of the population by training and test sets
| Variables | Training set | Test set | P |
|---|---|---|---|
| Number | 390 | 195 | |
| Age in years | 0.444 | ||
| Mean ± SD | 57.4 ± 8.4 | 56.8 ± 8.7 | |
| BMI (kg/m 2 ) | 0.096 | ||
| Mean ± SD | 24.2 ± 3.2 | 23.8 ± 3.3 | |
| NSE (ng/ml) | 0.126 | ||
| Median (Q1 - Q3) | 12.3–16.4 | 12.4–15.9 | |
| CEA (ng/ml) | 0.185 | ||
| Median (Q1 - Q3) | 1.4–4.3 | 1.3–3.9 | |
| CYFRA21-1 (ng/ml) | 0.302 | ||
| Median (Q1 - Q3) | 2.0–3.5 | 1.9–3.7 | |
| SCC (ng/ml) | 0.181 | ||
| Median (Q1 - Q3) | 0.6–1.1 | 0.6–1.2 | |
| Mediastinal fat area(cm 2 ) | 0.408 | ||
| Mean ± SD | 6.1 ± 3.0 | 5.9 ± 3.0 | |
| Sex | 0.230 | ||
| Female | 196 (50.3%) | 109 (55.9%) | |
| Male | 194 (49.7%) | 86 (44.1%) | |
| Smoking history | 0.929 | ||
| Yes | 167 (42.8%) | 85 (43.6%) | |
| No | 223 (57.2%) | 110 (56.4%) | |
| T stage | 0.051 | ||
| T1 + T2 | 355 (91.0%) | 187 (95.9%) | |
| T3 + T4 | 35 (9.0%) | 8 (4.1%) | |
| N stage | 0.414 | ||
| N0 | 276 (70.8%) | 146 (74.9%) | |
| N1 | 50 (12.8%) | 18 (9.2%) | |
| N2 | 64 (16.4%) | 31 (15.9%) | |
| ECOG performance status | 0.749 | ||
| 0 | 235 (60.3%) | 123 (63.1%) | |
| 1 | 135 (34.6%) | 64 (32.8%) | |
| 2 | 20 (5.1%) | 8 (4.1%) | |
| EGFR mutation | 0.568 | ||
| Positive | 102 (26.2%) | 46 (23.6%) | |
| Negative | 288 (73.8%) | 149 (76.4%) | |
| ALK mutation | 0.900 | ||
| Positive | 6 (1.5%) | 2 (1.0%) | |
| Negative | 384 (98.5%) | 193 (99.0%) | |
| Pathological type | 0.855 | ||
| Adenocarcinoma | 314 (80.5%) | 154 (79.0%) | |
| Squamous cell carcinoma | 67 (17.2%) | 37 (19.0%) | |
| Others | 9 (2.3%) | 4 (2.0%) | |
| Brain metastasis | 0.219 | ||
| Yes | 80 (20.5%) | 31 (15.9%) | |
| No | 310 (79.5%) | 164 (84.1%) |
BMI, body mass index; CYFRA21-1, cytokeratin 19 fragment 21-1; NSE, neuron-specific enolase; CEA, carcinoembryonic antigen; SCC, squamous cell carcinoma; ECOG, eastern cooperative oncology group performance status; EGFR, epidermal growth factor receptor; ALK, anaplastic lymphoma kinase
Deep learning model performance
DenseNet-121 outperformed comparators in the test set (AUC: 0.844 [95% CI: 0.769–0.919]; accuracy: 76.0% [69.7–82.1]; Table S3). The structural diagram of DenseNet-121 is shown in Figure S3. Thus, we selected DenseNet-121 as the feature extractor for 3D VOIs in our analysis of CT images. The Grad-CAM of the model is shown in Figure S4.
Multivariate predictors of BM
Independent BM determinants included N stage (OR: 2.308 [1.125–4.735], P = 0.022), EGFR mutation (OR: 2.920 [1.316–6.482], P = 0.008), MFA (OR: 1.087 [1.012–1.165], P = 0.013), and DLBMS (OR: 768.419 [157.056–3759.591], P < 0.001; Table 2). The large OR for DLBMS reflects its scaling as a logit-transformed probability. Clinically, DLBMS operates within a narrower range (0.2–0.8), where each 0.1-unit increase corresponds to a ~ 3.2-fold higher BM odds (derived from nomogram risk-score conversion). The 768.419 OR represents a theoretical full-range shift (0 → 1), emphasizing the model’s sensitivity to extreme DLBMS values.
Table 2.
The predictors of brain metastases in patients with NSCLC
| Variables | Univariable analysis | Multivariable analysis | ||
|---|---|---|---|---|
| OR (95% CI) | P-value | OR (95% CI) | P-value | |
| Sex (female vs. male) | 1.051 (0.643–1.719) | 0.842 | ||
| Age (years) | 0.987 (0.959–1.016) | 0.362 | ||
| Smoking status (yes vs. no) | 0.922 (0.560–1.518) | 0.750 | ||
| BMI (kg/m2) | 0.947 (0.876–1.025) | 0.177 | ||
| T stage (T3 + T4 vs. T1 + T2) | 2.919 (1.411–6.041) | 0.004 | 2.300 (0.829–6.385) | 0.110 |
| N stage (N1 + N2 vs. N0) | 8.013 (4.663–13.770) | < 0.001 | 2.308 (1.125–4.735) | 0.022 |
| ECOG performance status (0 vs. others) | 0.830 (0.499–1.382) | 0.474 | ||
| EGFR mutation (Positive vs. Negative) | 1.846 (1.091–3.124) | 0.022 | 2.920 (1.316–6.482) | 0.008 |
| ALK (Positive vs. Negative) | 3.987 (0.789–20.140) | 0.094 | ||
| Pathological type (Adenocarcinoma vs. Others) | 0.601 (0.301–1.201) | 0.149 | ||
| NSE (ng/ml) | 1.038 (0.991–1.086) | 0.114 | ||
| CEA (ng/ml) | 1.012 (1.002–1.022) | 0.024 | 1.005 (0.999–1.011) | 0.122 |
| CYFRA21-1 (ng/ml) | 1.032 (0.985–1.080) | 0.182 | ||
| SCC (ng/ml) | 0.964 (0.863–1.078) | 0.521 | ||
| Mediastinal fat area (cm2) | 1.110 (1.029–1.199) | 0.007 | 1.087 (1.012–1.165) | 0.013 |
| DLBMS |
1084.251 (258.644-4545.242) |
< 0.001 |
768.419 (157.056-3759.591) |
< 0.001 |
BMI, body mass index; CYFRA21-1, cytokeratin 19 fragment 21-1; NSE, neuron-specific enolase; CEA, carcinoembryonic antigen; SCC, squamous cell carcinoma; ECOG, eastern cooperative oncology group performance status; EGFR, epidermal growth factor receptor; ALK, anaplastic lymphoma kinase; DLBMS, deep learning brain metastasis score
Nomogram performance
Based on the findings from multivariate logistic regression analysis, a multimodal model was established and expressed as a nomogram (Fig. 2). The integrated nomogram demonstrated superior discrimination (test AUC: 0.947 [0.909–0.985]) for preoperative prediction of postoperative BM versus all comparators (P < 0.05; Table 3; Fig. 3). As shown in the probability distribution plot in Figure S5, the nomogram demonstrates an increasing number of cases of BM as the predicted probability increases in the test set. The confusion matrices used by the different models for predicting cases were depicted in Figure S6. Moreover, the nomogram could correct 83.3% positive BM cases and 89.3% negative cases in patients incorrectly diagnosed by the model combining MFA and clinical variables (Figure S7A). The nomogram prediction and the actual BM probability agreed well. According to the calibration curve analysis, the nomogram demonstrated significantly better predictive performance than other comparative models (Fig. 4A). DCA confirmed its preoperative clinical utility across risk thresholds (Fig. 4B). As depicted in Table 4, when comparing the nomogram with other models, there is a significant improvement in IDI (all P < 0.05) and NRI (all P < 0.05) for BM prediction.
Fig. 2.
The nomogram predicts brain metastases
Table 3.
The performance comparison of different models
| Models | AUC | ACC(%) | SENS(%) | SPEC(%) | |
|---|---|---|---|---|---|
| MFA + Clinical + DL | T | 0.954[0.827,0.979] | 93.0[89.7,95.8] | 87.5[78.0,94.9] | 94.3[91.2,97.2] |
| V | 0.949[0.902,0.996] | 92.0[82.3,99.1] | 84.2[72.1,97.2] | 94.0[85.2,99.7] | |
| I-T | 0.947[0.909,0.985] | 92.8[88.7,96.4] | 83.5[70.2,94.8] | 94.5[91.1,97.6] | |
| MFA + Clinical | T | 0.786[0.743,0.830] | 75.4[70.5,80.1] | 67.5[56.5,78.3] | 77.4[72.3,82.4] |
| V | 0.780[0.671,0.889] | 78.0[69.2,87.2] | 70.9[57.0,83.5] | 77.3[67.2,87.7] | |
| I-T | 0.783[0.691,0.875] | 79.5[73.8,85.1] | 61.6[44.4,77.8] | 82.9[77.2,88.3] | |
| MFA + DL | T | 0.910[0.880,0.940] | 79.5[75.0,84.0] | 93.6[85.6,98.1] | 75.4[70.2,80.8] |
| V | 0.885[0.784,0.985] | 78.2[69.2,87.2] | 81.3[65.1,95.2] | 77.4[66.7,87.3] | |
| I-T | 0.882[0.812,0.952] | 79.0[73.3,84.6] | 74.2[57.1,89.2] | 79.9[74.0,86.0] | |
| Clinical | T | 0.779[0.726,0.832] | 71.9[67.3,76.6] | 70.3[59.0,81.6] | 72.4[67.1,77.5] |
| V | 0.721[0.601,0.842] | 68.8[57.4,79.2] | 66.7[52.8,82.4] | 69.4[57.6,80.7] | |
| I-T | 0.709[0.612,0.806] | 68.2[61.0,74.9] | 61.3[44.8,78.9] | 69.4[62.6,76.4] | |
| MFA | T | 0.711[0.652,0.770] | 65.5[60.2,69.8] | 66.1[55.2,78.4] | 63.3[59.8,67.5] |
| V | 0.679[0.548,0.810] | 61.2[50.4,71.6] | 60.3[45.2,75.7] | 61.0[49.8,71.6] | |
| I-T | 0.704[0.590,0.818] | 62.6[55.9,69.2] | 64.5[48.0,80.8] | 62.2[55.3,69.3] | |
| DL | T | 0.912[0.883,0.941] | 82.5[78.5,86.5] | 92.2[85.5,98.2] | 80.0[75.0,84.7] |
| V | 0.884[0.785,0.982] | 78.5[67.9,87.2] | 82.0[68.5,94.0] | 77.5[66.7,87.5] | |
| I-T | 0.844[0.769,0.919] | 76.0[69.7,82.1] | 71.2[54.3,87.5] | 76.9[70.3,83.4] |
DL, deep learning; MFA, mediastinal fat area; AUC, area under the ROC curve; ACC, accuracy; SENS, sensitivity; SPEC, specificity; T, training set; V, validation set; I-T, independent test set; 95% confidence intervals included in brackets
Fig. 3.
Receiver operating characteristic (ROC) curves for comparison among various prediction models
Fig. 4.
(A) Calibration curves comparing different models. (B) Decision curve analysis comparing different models
Table 4.
NRI and IDI compared between different models
| Models | IDI | P | NRI | P |
|---|---|---|---|---|
| MFA + Clinical + DL model vs. MFA + Clinical model | 0.444 (0.335–0.553) | < 0.001 | 1.535 (1.281–1.788) | < 0.001 |
| MFA + Clinical + DL model vs. DL model | 0.301 (0.195–0.407) | < 0.001 | 0.616 (0.316–0.917) | < 0.001 |
| MFA + Clinical model vs. Clinical model | 0.123 (0.058–0.188) | < 0.001 | 0.386 (0.053–0.718) | 0.023 |
IDI, integrated discrimination improvement; NRI, net reclassification improvement; DL, deep learning; MFA, mediastinal fat area
MFA contribution
Incorporating preoperative MFA corrected 22.2% of false-negative and 23.7% of false-positive DL predictions (Figure S7B). MFA + clinical models outperformed clinical-only models (IDI: 0.123 [0.058–0.188]; NRI: 0.386 [0.053–0.718]; P < 0.05) (Table 4).
Survival prediction
Patients stratified as high-risk by the preoperative nomogram exhibited significantly reduced survival (log-rank P = 8.596 × 10⁻⁷; Fig. 5).
Fig. 5.
Kaplan-Meier survival analysis
Discussion
This study developed a nomogram integrating preoperative MFA and DL-derived tumor radiomics to predict postoperative BM risk and survival outcomes in resected NSCLC patients. To our knowledge, this represents the first investigation combining DL-based tumor characterization with adipose tissue metrics for preoperative prediction of postoperative BM risk stratification.
Prior studies have primarily relied on two-dimensional (2D) radiomic features or PET-CT imaging to predict metastasis [18–20]. While Chen et al. achieved an AUC of 0.847 for synchronous BM prediction using CT-derived 2D features in T1 lung adenocarcinoma, their model lacked external validation [19]. Similarly, PET-CT-based approaches, though effective, impose prohibitive costs for routine clinical use [20]. Our methodology advances this field by leveraging widely available preoperative chest CT scans to extract 3D tumor signatures via DL, capturing comprehensive morphological data while minimizing patient burden.
In addition, several clinical variables have been shown to be associated with BM [21, 22]. Our nomogram showed that pathological N stage and EGFR could predict the possibility of BM after surgery, which was consistent with most previous studies [21, 23, 24]. Notably, our research identified preoperative high MFA as an independent risk factor for postoperative BM. A lifetime of exposure to excessive amounts of adipose can enhance the risk of lung tumors [25]. Increased adipose tissue activity and volume are associated with worse cancer stage [26]. These results suggest that increased preoperative adipose tissue is beneficial for cancer development, which is consistent with our findings.
Fat may influence the occurrence of postoperative BM after NSCLC surgery through a variety of mechanisms [11]. Firstly, adipose tissue-induced inflammation and adipokine secretion play a key role in tumor progression [27]. Enlarged adipocytes experience hypoxia, leading to a chronic inflammatory state in the cells within adipose tissue. These cells release various pro-inflammatory cytokines, including tumor necrosis factor-alpha (TNF-α), interleukin-6 (IL-6), and interleukin-1 beta (IL-1β) [27]. The increase in these cytokines not only directly promotes the proliferation and survival of tumor cells but also activates related signaling pathways, such as NF-κB and JAK/STAT, further driving cancer development [27–29]. At the same time, adipokines such as leptin and adiponectin can regulate cell proliferation and immune escape, further contributing to the malignant progression of tumors [10]. Secondly, the imbalance of fat metabolism may drive the change of tumor microenvironment (TME) and increase the risk of metastasis [30]. Adipose-derived stem cells (ASCs) and cytokines can prevent host immune system from identifying and attacking cancer cells [31]. They may alter the immune balance in the TME, thereby suppressing anti-tumor immune responses [32]. In a mouse study, ASCs were observed to promote tumor growth by migrating from distant sites to the tumor microenvironment through the systemic circulation [33]. Fat alters the metabolic landscape of the TME, which inhibits T-cell function and promotes tumor growth [34]. Moreover, adipocytes remodel the extracellular matrix, supporting tumor growth and creating pathways for metastasis [35]. The lipid-rich environment provides energy to tumor cells, enhancing their metastatic potential [30]. Finally, adipocytes could regulate angiogenesis by secreting growth factors, and this process is essential for metastatic growth [27]. Adipocytes can actively release significant quantities of vascular endothelial growth factor (VEGF) [36]. VEGF is a potent pro-angiogenic factor that promotes the proliferation, migration, and formation of new blood vessels by binding to VEGF receptors on vascular endothelial cells, thereby promoting the rapid growth and metastasis of tumors [37]. These preclinical findings support our clinical observation that preoperative mediastinal adipose metrics predict postoperative metastasis risk.
Our study has several limitations that warrant discussion. First, the “black-box” nature of DL models inherently limits interpretability, as these systems cannot be easily decomposed into intuitive components for clinical understanding. Further validation with histological data or feature importance quantification would strengthen interpretability. Second, while our internal validation demonstrated robust performance, our dataset was derived from a single institution. This single-center design raises potential concerns about selection bias and may limit generalizability. External validation using multicenter datasets would strengthen confidence in the model’s clinical applicability across diverse patient populations and practice settings. Third, the retrospective design, while efficient for preliminary investigation, introduces inherent limitations in data quality control. Prospective studies are needed to verify the model’s predictive accuracy in real-world clinical workflows and to assess its impact on clinical decision-making. These steps will be crucial for translating our findings into practice.
Conclusion
The DL-enhanced nomogram, incorporating preoperative MFA, tumor radiomics, and clinicopathological variables, provides a robust tool for preoperative prediction of postoperative BM risk stratification in NSCLC.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Acknowledgements
Not applicable.
Author contributions
Y.N. conceived the study, performed data analysis, and wrote the original draft. H.B.J. contributed to methodology, software, and validation. X.M.L. contributed to methodology, software. W.J.H. handled data curation and analysis. P.P.L. was responsible for investigation and data curation. L.L. contributed to writing—review and editing. Z.Y.L. supervised the project and provided resources. Q.J.W. contributed to validation and analysis. Y.Z.L. assisted with software and validation. S.D.M. contributed to visualization and investigation. R.T.W. supervised the study, acquired funding, and contributed to writing—review and editing. Z.X.D. conceptualized the study and contributed to writing—review and editing.
Funding
This work was supported by Climbing program of Harbin Medical University Cancer Hospital (PDTS2024B-01).
Data availability
The data used in this study were collected in real world healthcare settings and access to these data is restricted due to privacy and proprietary reasons. Please contact the corresponding author for data and code access.
Declarations
Ethics approval
This study was conducted in accordance with the Declaration of Helsinki. The Ethical Review Committee of Harbin Medical University Cancer Hospital approved this study (approval number: YD2024–06).
Consent to participate
The need for informed consent was waived for this study by the Ethical Review Committee of Harbin Medical University Cancer Hospital, due to its retrospective design and irreversible anonymization of all data.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Ye Niu, Hao-Bo Jia, Xue-Meng Li authors have contributed equally to this work and share first authorship.
Contributor Information
Rui-Tao Wang, Email: ruitaowang@126.com.
Ze-Xun Duan, Email: duanzexun@sohu.com.
References
- 1.Zhang Q, et al. Risk factors of brain metastases as initial failure in completely resected stage IIIA(N2) non-small cell lung cancer. Ann Transl Med. 2020;8(6):374. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Ettinger DS, et al. Non-Small cell lung cancer, version 5.2017, NCCN clinical practice guidelines in oncology. J Natl Compr Canc Netw. 2017;15(4):504–35. [DOI] [PubMed] [Google Scholar]
- 3.Kim G, Moon S, Choi JH. Deep learning with multimodal integration for predicting recurrence in patients with Non-Small cell lung Cancer. Sens (Basel), 2022. 22(17). [DOI] [PMC free article] [PubMed]
- 4.Chalkidou A, O’Doherty MJ, Marsden PK. False discovery rates in PET and CT studies with texture features: A systematic review. PLoS ONE. 2015;10(5):e0124165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Zhong Y, et al. Deep learning for prediction of N2 metastasis and survival for clinical stage I non-small cell lung cancer. Radiology. 2022;302(1):200–11. [DOI] [PubMed] [Google Scholar]
- 6.Li X, et al. Diagnosis of thyroid cancer using deep convolutional neural network models applied to sonographic images: a retrospective, multicohort, diagnostic study. Lancet Oncol. 2019;20(2):193–201. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Bi WL, et al. Artificial intelligence in cancer imaging: clinical challenges and applications. CA Cancer J Clin. 2019;69(2):127–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Xu Y, et al. Deep learning predicts lung cancer treatment response from serial medical imaging. Clin Cancer Res. 2019;25(11):3266–75. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Liu J, et al. Associations of muscle and adipose tissue parameters with long-term outcomes in middle and low rectal cancer: a retrospective cohort study. Cancer Imaging. 2023;23(1):5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Dumas JF, Brisson L. Interaction between adipose tissue and cancer cells: role for cancer progression. Cancer Metastasis Rev. 2021;40(1):31–46. [DOI] [PubMed] [Google Scholar]
- 11.Brown KA, Scherer PE. Update on adipose tissue and Cancer. Endocr Rev. 2023;44(6):961–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zuo Z, et al. Survival nomogram for stage IB Non-Small-Cell lung Cancer patients, based on the SEER database and an external validation cohort. Ann Surg Oncol. 2021;28(7):3941–50. [DOI] [PubMed] [Google Scholar]
- 13.Byrne CA et al. Validation of skeletal muscle and adipose tissue measurements using a fully automated body composition analysis neural network versus a semi-automatic reference program with human correction in patients with lung cancer. Heliyon, 2022. 8(12). [DOI] [PMC free article] [PubMed]
- 14.Ponholzer F et al. Sarcopenia and mediastinal adipose tissue as a prognostic marker for short- and long-term outcomes after primary surgical treatment for lung cancer. Cancers (Basel), 2023. 15(23). [DOI] [PMC free article] [PubMed]
- 15.Chen O, et al. Correlation between pericardial, mediastinal, and intrathoracic fat volumes with the presence and severity of coronary artery disease, metabolic syndrome, and cardiac risk factors. Eur Heart J Cardiovasc Imaging. 2015;16(1):37–46. [DOI] [PubMed] [Google Scholar]
- 16.Rozenberg D, et al. Clinical outcomes associated with computed tomography-based body composition measures in lung transplantation: a systematic review. Transpl Int. 2020;33(12):1610–25. [DOI] [PubMed] [Google Scholar]
- 17.Kwack WG, et al. Association between thoracic fat measured using computed tomography and lung function in a population without respiratory diseases. J Thorac Disease. 2019;11(12):5300–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Liu X, et al. Deep learning radiomics-based prediction of distant metastasis in patients with locally advanced rectal cancer after neoadjuvant chemoradiotherapy: A multicentre study. EBioMedicine. 2021;69:103442. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Chen A, et al. CT-Based radiomics model for predicting brain metastasis in category T1 lung adenocarcinoma. AJR Am J Roentgenol. 2019;213(1):134–9. [DOI] [PubMed] [Google Scholar]
- 20.Zheng Z, et al. (18)F-FDG PET/CT radiomics predicts brain metastasis in I-IIIA resected Non-Small cell lung cancer. Eur J Radiol. 2023;165:110933. [DOI] [PubMed] [Google Scholar]
- 21.Won YW, et al. A nomogram to predict brain metastasis as the first relapse in curatively resected non-small cell lung cancer patients. Lung Cancer. 2015;88(2):201–7. [DOI] [PubMed] [Google Scholar]
- 22.Sun S, et al. A nomogram for predicting brain metastasis in IIIA-N2 non-small cell lung cancer after complete resection: a competing risk analysis. Front Oncol. 2021;11:781340. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Wang H, et al. Driver genes as predictive indicators of brain metastasis in patients with advanced NSCLC: EGFR, ALK, and RET gene mutations. Cancer Med. 2020;9(2):487–95. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Ge M, et al. High probability and frequency of EGFR mutations in non-small cell lung cancer with brain metastases. J Neurooncol. 2017;135(2):413–8. [DOI] [PubMed] [Google Scholar]
- 25.Lu Y, et al. Assessment of causal effects of visceral adipose tissue on risk of cancers: a Mendelian randomization study. Int J Epidemiol. 2022;51(4):1204–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Chu K, et al. Brown adipose tissue and cancer progression. Skeletal Radiol. 2019;49(4):635–9. [DOI] [PubMed] [Google Scholar]
- 27.Cozzo AJ, Fuller AM, Makowski L. Contribution of adipose tissue to development of Cancer. Compr Physiol. 2017;8(1):237–82. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Lengyel E, et al. Cancer as a matter of fat: the crosstalk between adipose tissue and tumors. Trends Cancer. 2018;4(5):374–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Yadav A, et al. IL-6 promotes head and neck tumor metastasis by inducing epithelial-mesenchymal transition via the JAK-STAT3-SNAIL signaling pathway. Mol Cancer Res. 2011;9(12):1658–67. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Cozzo AJ, Fuller AM, Makowski L. Contribution of Adipose Tissue to Development of Cancer, in Comprehensive Physiology. 2017. pp. 237–282. [DOI] [PMC free article] [PubMed]
- 31.Ko SF, et al. Adipose-Derived mesenchymal stem cell exosomes suppress hepatocellular carcinoma growth in a rat model: apparent diffusion coefficient, natural killer T-Cell responses, and histopathological features. Stem Cells Int. 2015;2015:853506. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Iyengar NM, et al. Obesity and cancer mechanisms: tumor microenvironment and inflammation. J Clin Oncol. 2016;34(35):4270–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Scioli MG et al. Adipose-Derived stem cells in Cancer progression: new perspectives and opportunities. Int J Mol Sci, 2019. 20(13). [DOI] [PMC free article] [PubMed]
- 34.Ringel AE, et al. Obesity shapes metabolism in the tumor microenvironment to suppress Anti-Tumor immunity. Cell. 2020;183(7):1848–e186626. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Seo BR, et al. Obesity-dependent changes in interstitial ECM mechanics promote breast tumorigenesis. Sci Transl Med. 2015;7(301):301ra130. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Nijhawans P, Behl T, Bhardwaj S. Angiogenesis in obesity. Biomed Pharmacother. 2020;126:110103. [DOI] [PubMed] [Google Scholar]
- 37.Skrypnik D, et al. Leptin-VEGF crosstalk in excess body mass and related disorders: A systematic review. Obes Rev. 2023;24(8):e13575. [DOI] [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 data used in this study were collected in real world healthcare settings and access to these data is restricted due to privacy and proprietary reasons. Please contact the corresponding author for data and code access.