TL;DR
This study develops a novel nomogram using CT-based radiomics and clinical features to predict "Spread Through Air Spaces" (STAS) in early-stage non-small cell lung cancer (NSCLC). The proposed model significantly outperforms traditional clinical assessments, offering a non-invasive tool to guide surgical decisions and potentially improve patient outcomes.
What Problem Does This Paper Address?
Spread Through Air Spaces (STAS) is an aggressive invasion pattern in non-small cell lung cancer (NSCLC) that is associated with a higher risk of recurrence and poorer prognosis. Currently, the definitive diagnosis of STAS relies on postoperative histopathological examination, which is invasive and does not inform preoperative surgical planning. While CT imaging has shown promise in identifying STAS, existing radiomics models have limitations in their generalizability and predictive accuracy. This paper aims to address this gap by developing a robust, multi-center predictive model that integrates both intratumoral and peritumoral radiomics features with clinical and radiological indicators for early-stage NSCLC (Stage IA).
Methods
The researchers employed a multi-stage approach to build and validate their predictive model.
- Data Collection and Cohort Splitting: Data from 405 NSCLC patients were collected from three hospitals. These were divided into a training cohort (n=236), an internal validation cohort (n=100), and an external validation cohort (n=69). Patients with pure ground-glass nodules (pGGNs), which are known to be STAS-negative, were excluded.
- Feature Extraction:
- Clinical and Radiological Features: Standard clinical data (age, sex, smoking status, etc.) and qualitative radiological features (tumor location, density type, spiculation, etc.) were extracted and analyzed.
- Radiomics Features: The core of the method involved extracting radiomics features from CT images. This included:
- Gross Tumor Volume (GTV): Features from the tumor itself.
- Peritumoral Volume (PTV): Features from regions extending 5, 10, 15, and 20 mm around the tumor.
-
Gross Peritumoral and Tumor Volume (GPTV): Features from combined intratumoral and peritumoral regions. A total of 1218 radiomics features were extracted, encompassing first-order statistics, shape, and texture features.
-
Model Development:
- Radiomics Models: Nine radiomics models were constructed based on different volumes (GTV, PTVs, GPTVs). Feature selection was performed using the mRMR algorithm and LASSO logistic regression to mitigate overfitting and identify the most predictive features.
- Clinical Model: A clinical model was built using univariate and multivariate logistic regression analysis to identify independent risk factors from clinical and radiological features.
- Combined Model: The best-performing radiomics model (GPTV10) and the selected clinical-radiological predictors were combined to create a comprehensive predictive model.
- Model Evaluation: The performance of all models was evaluated using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values across the three cohorts. DeLong tests were used to compare the AUC values between different models. Hosmer-Lemeshow tests and calibration curves assessed model fit, and decision curve analysis (DCA) evaluated clinical utility.
Datasets
The study utilized retrospective data from three hospitals, comprising a total of 405 NSCLC patients.
- Hospital 1: 336 patients (training cohort: n=236, internal validation cohort: n=100).
- Hospital 2: 30 patients (external validation cohort).
- Hospital 3: 39 patients (external validation cohort). The combined external validation cohort consisted of 69 patients. No new benchmark datasets were introduced; the study focused on validating its model across existing patient data.
Key Insights and Ideas
- Peritumoral Radiomics Matters: The study highlights that not only the tumor itself (GTV) but also the surrounding peritumoral region (PTV and GPTV) contains crucial information for predicting STAS. The GPTV models, which incorporate both tumor and peritumoral information, generally outperformed PTV models.
- The "Sweet Spot" for Peritumoral Expansion: The GPTV10 model (incorporating a 10mm peritumoral expansion) emerged as the best radiomics model, suggesting an optimal range for capturing relevant peritumoral features for STAS prediction. This aligns with previous research indicating that STAS is often found within a certain distance from the primary tumor.
- Clinical and Radiomics Synergy: Combining the best radiomics model (GPTV10) with key clinical and radiological predictors (density type, distal ribbon sign) resulted in a significantly superior predictive model compared to either approach alone. This underscores the value of integrating diverse data sources.
- Radiomics Overcomes Subjectivity: The paper emphasizes that radiomics can extract objective, quantitative features from CT images, overcoming the inherent subjectivity of traditional radiological assessments by experienced radiologists.
Key Results
- Independent Risk Factors: Tumor density type (OR=6.738) and distal ribbon sign (OR=5.141) were identified as independent risk factors for STAS.
- Best Radiomics Model: The GPTV10 radiomics model demonstrated strong predictive performance with AUC values of 0.887 (training), 0.876 (internal validation), and 0.868 (external validation).
- Clinical Model Performance: The clinical model achieved AUC values of 0.874 (training), 0.822 (internal validation), and 0.810 (external validation).
- Combined Model Superiority: The combined model, integrating GPTV10 radscore and clinical predictors, achieved the highest AUC values: 0.901 (training), 0.875 (internal validation), and 0.878 (external validation). DeLong tests confirmed the combined model's superiority over the clinical model in all cohorts.
- Nomogram Utility: A nomogram was developed based on the combined model, providing an intuitive visual tool for predicting STAS risk, which was shown to have better clinical application value than the individual models.
Limitations and Future Work
The authors acknowledge several limitations:
- Sample Size: While efforts were made to select representative cases, the overall sample size was relatively small.
- Retrospective Design: The retrospective nature of the study introduces potential selection bias.
- Lack of Follow-up Data: The current study did not include detailed follow-up data, limiting its ability to predict patient prognosis or perform risk stratification.
Future work will involve:
- Prospective Studies: Conducting prospective studies with larger, multi-center datasets to further validate the nomogram and improve its generalizability.
- Deep Learning Integration: Exploring the integration of deep learning techniques, which have shown promise in similar tasks, to potentially enhance prediction performance.
- Prognostic Prediction: Investigating the nomogram's utility in predicting patient prognosis and risk stratification once sufficient follow-up data are collected.
Summary updated