Construction of novel radiomics nomogram model based on preoperative CT to predict lymphovascular tumor embolus and recurrence-free survival in early T1-2a stage lung adenocarcinomas

BMC Med Imaging 2026 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Page 2
Lymphovascular Invasion in Early Lung Cancer

The Recurrence Problem. Even with surgical resection of early-stage non-small cell lung cancer (NSCLC), outcomes remain inadequate. Patients with stage I NSCLC have a 33% recurrence rate and only a 51% five-year survival rate, making preoperative identification of high-risk features essential for treatment planning.

What is Lymphovascular Tumor Embolus. Lymphovascular tumor embolus (LTE) occurs when malignant cells invade arteries, veins, or lymphatic vessels and form cancer thrombi. This vascular invasion is a critical mechanism of metastasis and is strongly associated with disease recurrence and death in lung adenocarcinoma patients.

The Detection Challenge. LTE is a histologically defined condition that can only be confirmed after surgery by examining resected tissue specimens. Conventional CT imaging can detect large macrovascular thrombi but cannot identify microscopic vascular invasion (MVI) or LTE preoperatively.

Why Preoperative Prediction Matters. NCCN guidelines designate MVI as a high-risk factor, recommending adjuvant chemotherapy for stage IB NSCLC patients with this finding. If LTE could be predicted before surgery using imaging, treatment strategies could be adjusted proactively rather than reactively.

TL;DR: Lymphovascular tumor embolus predicts recurrence in early lung adenocarcinoma but can only be confirmed after surgery, motivating the need for preoperative CT-based prediction tools.
Pages 2-4
Radiomics Feature Extraction and Model Design

Patient Population. The study enrolled 195 patients with pathologically confirmed lung invasive adenocarcinoma (LAC) treated at Weifang People's Hospital between January 2018 and April 2022. Of these, 43 had LTE and 152 did not. The dataset was split into a training cohort of 144 patients and a validation cohort of 51 patients.

CT Imaging Protocol. All patients underwent preoperative chest CT on a 64-slice spiral scanner with standardized acquisition parameters (120 kV, 1.25 mm reconstruction thickness). Two radiologists with 7 and 16 years of thoracic imaging experience manually delineated tumor regions of interest (ROIs) using ITK-SNAP software, blinded to clinical information.

Feature Extraction. Using the Pyradiomics software package, 850 radiomic features were extracted from each CT scan. These included first-order statistics, shape features (2D and 3D), and multiple gray-level texture matrices including the gray level co-occurrence matrix, gray level run length matrix, and wavelet-transformed features.

Feature Selection. LASSO (Least Absolute Shrinkage and Selection Operator) regression with 10-fold cross-validation reduced 850 features to just 5 for model construction. This dimensionality reduction prevents overfitting and identifies the most predictive imaging biomarkers from a large feature pool.

Model Types. Two predictive models were built: a pure radiomics model based solely on CT texture features, and a hybrid combined model that also incorporated smoking status as an independent clinical predictor. Both were evaluated using AUC, calibration curves, and decision curve analysis.

TL;DR: From 850 CT radiomics features, LASSO regression selected 5 key features to build two LTE prediction models -- one pure radiomics and one hybrid including smoking status.
Pages 5-6
Model Performance and Key Predictors

The Five Selected Features. LASSO regression retained five radiomic features for the radiomics score (RS): LongRunHighGrayLevelEmphasis, Autocorrelation, Idn, DependenceNonUniformityNormalized, and Skewness. These features capture tumor texture heterogeneity and gray-level distribution patterns within the tumor.

Smoking as a Risk Factor. Univariate analysis identified smoking status as an independent predictor of LTE, with an odds ratio of 2.89. LTE was present in 60.5% of smokers but only 39.5% of non-smokers in the LTE-positive group. Smoking appears to remodel tumor vasculature through HIF-1a/VEGF axis dysregulation, generating leaky vessels that facilitate tumor emboli formation.

AUC Performance. In the training cohort, the combined model achieved an AUC of 0.921 versus 0.881 for the radiomics-only model. In the validation cohort, both models performed similarly (AUC of 0.878 and 0.877, respectively). The combined model showed better sensitivity (83.3% vs 77.4%) and specificity (89.7% vs 88.5%) in the training cohort.

Calibration and Clinical Utility. Calibration curves demonstrated good agreement between predicted and actual LTE probabilities in both cohorts. Decision curve analysis confirmed that both models provided net clinical benefit across a range of threshold probabilities, supporting their potential utility in clinical decision-making.

TL;DR: The hybrid model combining CT radiomics and smoking status achieved an AUC of 0.921 in training and 0.878 in validation, outperforming the radiomics-only model.
Pages 5, 7
Survival Analysis Findings

Recurrence-Free Survival Outcomes. Among 195 patients followed for at least five years, the median recurrence-free survival (RFS) was 24 months, with 1-, 2-, and 3-year RFS rates of 83.9%, 80.9%, and 61.1% respectively. These figures underscore the continued recurrence risk even in early-stage disease.

Risk Group Stratification. Kaplan-Meier analysis showed a statistically significant difference in RFS between patients classified as low-score and high-score by the combined model (p < 0.05). This demonstrates that the model's radiomics-derived score can stratify patients into meaningfully different prognostic groups.

LTE as a Survival Predictor. Patients who were LTE-positive showed significantly worse recurrence-free survival than LTE-negative patients. This confirms the clinical relevance of preoperatively identifying LTE: it is not just a pathological marker but a direct indicator of inferior oncological outcomes.

Practical Implications. By stratifying patients into high-risk and low-risk groups preoperatively, the nomogram could guide decisions about adjuvant chemotherapy, closer surveillance intervals, and enrollment in clinical trials for patients with early-stage but biologically aggressive disease.

TL;DR: Patients classified as high-risk by the combined nomogram had significantly shorter recurrence-free survival, validating the model's prognostic value beyond simple LTE status prediction.
Pages 6-7
Radiomics Biomarkers and Clinical Context

Autocorrelation as a Shared Biomarker. One of the five selected features, Autocorrelation, was also identified in a prior study by Deng et al. predicting MVI in stage I NSCLC. The independent replication of this feature across different studies suggests it may be a robust and biologically meaningful marker of vascular invasion in lung cancer.

Intratumoral Focus. Unlike some studies that combine intratumoral and peritumoral radiomics features, this study focused solely on intratumoral regions. The authors justify this by noting that intratumoral features better reflect tumor proliferation and heterogeneity, and that peritumoral boundary definitions remain unstandardized across institutions.

Biological Mechanism of Smoking. The study provides a mechanistic explanation for smoking's role in LTE: smoke exposure dysregulates the HIF-1a/VEGF signaling axis, promoting angiogenesis and creating structurally abnormal, leaky blood vessels that facilitate cancer cell entrapment and embolus formation.

Comparison to Existing Research. Prior radiomics studies on MVI in liver cancers achieved AUCs ranging from 0.72 to 0.995. This study's combined model AUC of 0.921 in training places it favorably within this range, though external validation using public datasets such as TCIA would strengthen generalizability claims.

TL;DR: Autocorrelation emerged as a consistent biomarker for vascular invasion in lung cancer, and smoking status mechanistically promotes LTE through vascular remodeling via the HIF-1a/VEGF pathway.
Pages 7-8
Clinical Value and Study Limitations

Main Finding. A CT radiomics-based hybrid nomogram combining five imaging features and smoking status can effectively predict LTE status and recurrence-free survival in early T1-2a stage lung adenocarcinomas with high accuracy, providing a noninvasive preoperative tool.

Treatment Implications. Accurate preoperative identification of LTE can enable timely adjustment of treatment plans, such as recommending adjuvant chemotherapy, planning more comprehensive surgical resections, or intensifying postoperative surveillance for high-risk patients identified before the operation.

Study Limitations. The study was retrospective and limited to a single institution, with a relatively small sample of 195 patients. The absence of external validation using independent cohorts or public datasets (like TCIA) limits generalizability. Only early-stage T1-2a disease was included, excluding advanced-stage patients.

Future Directions. Prospective multicenter studies with larger and more diverse patient cohorts are needed to validate these findings. Future work could also explore whether adding peritumoral features, dynamic contrast-enhanced imaging, or molecular markers further improves LTE prediction and survival stratification in early lung adenocarcinoma.

TL;DR: A CT radiomics nomogram combining five imaging features and smoking status reliably predicts preoperative LTE status and survival risk in early-stage lung adenocarcinoma, pending multicenter validation.
Citation: Open Access, 2026. Available at: PMC13036914.