The clinical decision problem Ground-glass nodules (GGNs) on lung CT scans are a common finding and can represent either minimally invasive adenocarcinoma (MIA) or invasive adenocarcinoma (IAC). The distinction is critical: MIA can be treated with limited wedge resection while IAC often requires lobectomy. Predicting invasiveness before surgery avoids overtreatment of MIA and undertreatment of IAC.
Study approach Researchers at Benxi Central Hospital analyzed 323 ground-glass nodules from 317 patients. Rather than focusing only on features within the tumor itself, they extracted radiomic features from the peritumoral region - the tissue surrounding the nodule at 1, 2, 3, 4, and 5 mm margins. Eight different machine learning models were then compared for each margin distance.
Key findings All models with all margin distances achieved AUC values above 0.75, and optimal models (all except decision tree) achieved AUCs approximately 0.90 in training and 0.85-0.91 in validation. No single optimal margin distance was identified - all five performed comparably (no statistically significant differences by DeLong test), though smaller margins tended to perform slightly better in training.
Practical implication Peritumoral radiomic features - the tissue zone just outside the tumor - contain valuable information about tumor invasiveness, reflecting biologically meaningful interactions between the tumor and its surrounding microenvironment. This expands the radiomic analysis beyond the tumor boundary into clinically useful territory.
Ground-glass nodules and their significance Ground-glass nodules (GGNs) appear as hazy, cloud-like opacities on CT scans that do not obscure underlying structures. Most lung adenocarcinomas detected on CT screening present as GGNs. The critical distinction is between MIA (less than 5mm stromal invasion) and IAC (greater than 5mm invasion or spread to lymphovascular structures).
Why surgery type matters MIA has a 5-year survival rate approaching 100% after segmental resection, while IAC has lower survival and requires more extensive surgery (lobectomy) to prevent recurrence. Unnecessarily performing lobectomy on an MIA patient removes more healthy lung tissue than needed, increasing morbidity. Conversely, performing only segmentectomy on an IAC risks inadequate margins and recurrence.
Current imaging limitations Radiologists assess GGN morphology (size, pure vs. mixed ground-glass, lobulation) to estimate invasiveness, but purely visual assessment has limited accuracy. Radiomics offers a way to extract quantitative features that correlate with histopathological invasiveness more precisely than subjective visual assessment.
The peritumoral biology rationale Microscopic studies have identified a 0-3.78mm transition zone between lung adenocarcinoma and normal lung tissue. The tumor-lung interface shows biological changes including angiogenesis, inflammatory cell infiltration, and stromal remodeling that are captured in CT texture patterns just outside the visible tumor boundary - the rationale for peritumoral radiomics.
Patient cohort 317 patients (239 female, 78 male; mean age 54.9 years) with 323 surgically confirmed pulmonary GGNs were enrolled from Benxi Central Hospital (January 2019 to December 2023). Inclusion required nodule diameter up to 30mm presenting as GGN, high-resolution CT within 1 month before surgery, and pathological confirmation as MIA or IAC. The cohort was 48.6% MIA and 51.4% IAC.
Peritumoral region segmentation A senior radiologist manually delineated each nodule in 3D using 3D Slicer software. The tumor contour was then expanded outward by 1, 2, 3, 4, and 5 mm to create five separate peritumoral regions of interest (ROIs). Normal structures (blood vessels, bronchioles, pleura, chest wall) within each expanded ROI were manually excluded to prevent contamination of peritumoral features.
Radiomic feature extraction 822 radiomic features were extracted from each peritumoral ROI: first-order statistical, shape, grey-level co-occurrence matrix, grey-level run matrix, grey-level dependence matrix, grey-level region scale matrix, neighborhood grey-level difference matrix, and wavelet transform features. MRMR and LASSO regression with 10-fold cross-validation reduced features to 6-16 per margin distance.
Eight machine learning models Logistic regression, adaptive boosting, random forest, Naive Bayes, decision tree, support vector machine, K-nearest neighbor, and neural network models were trained for each of the five margin distances. The 7:3 train/validation split was maintained, and DeLong tests were used for statistical comparisons between model AUC values.
Training set AUC values The best AUC for each margin distance in the training set was: 1mm 0.926, 2mm 0.925, 3mm 0.920, 4mm 0.924, 5mm 0.923. The ranking was 1mm greater than 2mm greater than 4mm greater than 5mm greater than 3mm. All models except decision tree achieved AUC approximately 0.90.
Validation set AUC values Best AUC in the validation set: 3mm 0.908, 2mm 0.903, 1mm 0.892, 4mm 0.892, 5mm 0.855. The ranking reversed partially compared to training. Sensitivity, specificity, accuracy, PPV, and NPV all exceeded 0.80 for all models in both training and validation, demonstrating consistently strong classification performance.
No statistically significant differences between margins DeLong tests comparing all five peritumoral margin models found no statistically significant differences in AUC in either training or validation sets (p greater than 0.05 for all pairwise comparisons). The decision tree was the only model significantly worse than all others, consistent with its lower AUC range of 0.75-0.85.
Clinical group differences IAC nodules were significantly larger (mean 12.4mm vs. 7.4mm in MIA, p less than 0.01), more often mixed ground-glass (54.2% vs. 17.8%, p less than 0.01), and occurred in older patients (57.6 vs. 52.1 years, p less than 0.01). No significant differences were found by sex or nodule location.
Pre-surgical decision support A validated model predicting MIA vs. IAC from pre-operative CT scans could inform the surgical planning discussion. Patients with model-predicted MIA could be counseled about limited resection options, while those predicted IAC could be prepared for lobectomy. This reduces operative surprise and supports more informed shared decision-making.
Smaller margins appear adequate Although no margin proved statistically superior, the trend suggests that the 1-3mm peritumoral range captures the most relevant information. This is biologically consistent with the known 0-3.78mm transition zone between adenocarcinoma and normal lung tissue - suggesting the relevant biological signals are captured within this narrow band.
Value beyond intratumoral features Prior studies using intratumoral radiomic models achieved similar AUC values (0.80-0.90). The peritumoral approach achieves comparable performance by capturing information from tissue that radiologists do not currently characterize. Future combined intratumoral and peritumoral models may outperform either approach alone.
Multiple ML models validated The finding that seven of eight machine learning models perform equivalently provides flexibility for clinical implementation - institutions can choose models that best fit their technical infrastructure and interpretability needs without sacrificing predictive accuracy. Avoiding the decision tree is the key practical takeaway from the model comparison.
Single-center retrospective design All patients were from Benxi Central Hospital, scanned on a single GE CT scanner under uniform acquisition protocols. While this standardization reduces imaging variability within the study, it limits generalizability to other institutions with different scanners and protocols. Single-center studies are known to overestimate model performance.
No external validation cohort The study used a single-center 7:3 train/validation split rather than an independent external validation cohort from a different institution. External validation with data from multiple hospitals using different scanner protocols is essential to establish the model's robustness in real-world diverse imaging environments.
Manual segmentation dependency All peritumoral ROIs were manually delineated by radiologists, which is time-consuming and subject to inter-observer variability. Automated or semi-automated segmentation approaches are needed to make this workflow clinically scalable. The study assessed intra-observer reproducibility for 30 nodules, but full inter-observer reliability was not reported.
Future directions Multicenter prospective validation studies with diverse CT scanners are needed. Development of automated peritumoral segmentation pipelines and combination models integrating intratumoral and peritumoral features could improve accuracy. Future studies should also include patients with atypical adenomatous hyperplasia and adenocarcinoma in situ to extend the classification to pre-invasive lesions.