Tumor spread through air spaces in non-small-cell lung cancer: 18F-fluorodeoxyglucose PET-based tumoral and peritumoral radiomics analysis with machine learning
Nur Aydinbelge Dizdar, Ebru Tatci, Aysu Dursun, Kayacan Dizdar, Özge Kaya Korkmaz, Ozlem OzmenObjective
To investigate the utility of tumoral and peritumoral [ 18 F]-fluorodeoxyglucose PET-based radiomics models for predicting tumor spread through air spaces (STAS) in non-small-cell lung cancer (NSCLC).
Methods
A total of 104 patients with NSCLC were retrospectively included and classified as STAS-positive or STAS-negative according to postoperative histopathology. Three radiomics models were developed based on tumoral and peritumoral volumes of interest (VOIs). Peritumoral models were constructed by defining VOIs with 5 and 10 mm expansions from the tumor margin (PR-5 and PR-10). Conventional PET-derived parameters were calculated. Each group was randomly divided into training (70%) and testing (30%) sets. For prediction, five machine learning algorithms were applied. Model performance was assessed by area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.
Results
Of the 104 patients, 60 (57.7%) were in the STAS-positive group, and 44 (42.3%) were in the STAS-negative group. Statistically significant differences were found in 27 tumoral, 62 PR-5, and 89 PR-10 radiomic features between the groups. Support Vector Machine achieved the best performance for the tumoral model (AUC: 0.70), whereas Logistic Regression was optimal for the PR-5 and PR-10 models, with AUCs of 0.79 and 0.74, respectively. The mean F1 score values for the tumoral, PR-5, and PR-10 models were 0.67, 0.72, and 0.68, respectively. Maximum standardized uptake value (SUV max ), SUV mean , and SUV peak did not differ significantly between groups. Tumor size, total lesion glycolysis, and metabolic tumor volume were significantly higher in the STAS-positive group.
Conclusion
Our findings suggest that peritumoral PET-based radiomics may help predict STAS status in NSCLC. The PR-5 model demonstrated the highest predictive performance, while conventional SUV-based parameters showed limited predictive value.