Clinical Challenge: Distinguishing small-cell lung cancer (SCLC) from non-small-cell lung cancer (NSCLC) preoperatively is critical because SCLC is primarily managed with chemotherapy and immunotherapy, not surgery, while early-stage NSCLC may be curable with resection.
Study Innovation: This study developed a CT radiomics nomogram incorporating peritumoral features - extracted from a 2mm expanded region surrounding the tumor - to non-invasively differentiate SCLC from NSCLC preoperatively.
Key Finding: The expanded peritumoral ROI (AUC 0.85) outperformed the standard intratumoral ROI (AUC 0.76) and clinical features alone (AUC 0.71), while a combined clinical-radiomics nomogram achieved AUC 0.96.
Study Population: 113 patients (54 SCLC, 59 NSCLC) with 1,050 radiomic features extracted, providing a focused dataset to validate the peritumoral radiomics concept.
Tumor Microenvironment: The tissue immediately surrounding a lung tumor reflects the microenvironment, including inflammatory infiltrates, desmoplastic stroma, and vascular invasion patterns that differ between SCLC and NSCLC.
2mm Expansion Method: The peritumoral ROI was created by expanding the manually segmented tumor contour outward by 2mm in 3D, capturing the peri-lesional tissue zone without requiring additional manual delineation.
Biological Rationale: SCLC's neuroendocrine biology and rapid growth characteristics create a distinct peri-tumoral reaction pattern compared to NSCLC, which may be captured by radiomic texture features even when intratumoral features overlap.
Added Information: The peritumoral zone captures vascular and lymphatic invasion patterns that extend beyond the visible tumor margin, providing information about biological aggressiveness not reflected in the tumor's internal imaging features alone.
Feature Extraction: 1,050 radiomic features were extracted from CT images using PyRadiomics, spanning first-order statistics, shape descriptors, and texture features (GLCM, GLRLM, GLSZM) from both original and wavelet-transformed images.
ROI Definitions: Three ROIs were compared: the original intratumoral ROI, the 2mm expanded peritumoral ROI, and a combination of both - with each analyzed independently to isolate the contribution of peri-tumoral versus intratumoral texture.
Feature Selection: LASSO regression with cross-validation was applied to select the most discriminative radiomic features from the 1,050 candidates, reducing dimensionality while preserving predictive information.
Nomogram Construction: A logistic regression model combining selected radiomic features with significant clinical variables (including CT morphological features and clinical characteristics) was visualized as a nomogram for clinical use.
Combined Nomogram: The clinical-radiomics nomogram achieved AUC 0.96 (95% CI: 0.88-1.00) and accuracy 0.91 in the validation cohort, representing the best-performing model.
Peritumoral Advantage: The peritumoral expanded ROI model (AUC 0.85) outperformed the original intratumoral ROI model (AUC 0.76), confirming that peritumoral texture information adds diagnostic value.
Clinical Model Baseline: Clinical features alone achieved AUC 0.71, providing a baseline that both radiomics approaches significantly exceeded.
Incremental Value: Each step - from clinical features to intratumoral radiomics to peritumoral radiomics to combined nomogram - provided meaningful incremental improvement, validating the hierarchical contribution of each information source.
Surgery vs. Chemotherapy Decision: For an apparent early-stage lung mass, accurately predicting SCLC preoperatively avoids unnecessary thoracotomy, as SCLC - even when localized - typically does not benefit from surgery in the same way NSCLC does.
Biopsy Challenges: Some central or small lung tumors are difficult to biopsy safely; in these cases, a high-confidence SCLC prediction from CT radiomics could support directing the patient to bronchoscopic or medical oncology rather than thoracic surgery.
Tumor Board Utility: The nomogram could be presented at multidisciplinary tumor board discussions to complement pathological results, particularly when biopsy samples are insufficient or inconclusive for definitive SCLC vs. NSCLC classification.
Risk Stratification: Patients predicted as SCLC but with ambiguous pathology could be prioritized for repeat biopsy or molecular profiling to confirm the diagnosis before committing to a treatment approach.
Small Sample Size: With only 113 patients, this study is underpowered for robust internal validation and may not capture the full spectrum of SCLC and NSCLC morphological presentations.
Single-Center Design: CT acquisition parameters, scanner type, and reconstruction algorithms at a single center may not reflect the variability encountered across institutions, limiting nomogram generalizability.
Manual Segmentation: Manual ROI delineation is time-consuming and subject to inter-observer variability; automating segmentation using deep learning would be essential for practical clinical deployment.
External Validation Needed: Validation in independent multicenter cohorts with diverse CT protocols is required to confirm that the peritumoral radiomics advantage holds broadly and that AUC 0.96 is not an artifact of the small, single-center training set.