Abstract
BACKGROUND: This study evaluates the effectiveness of radiomics based on 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) in distinguishing lung adenocarcinoma from squamous cell carcinoma, the 2 major subtypes of non-small cell lung cancer (NSCLC). METHODS: We extracted radiomic features from both the intratumoral and peritumoral areas in NSCLC patients' PET/CT scans. The intratumoral volume of interest (VOI_I) was manually outlined on PET/CT images. Additionally, 4 peritumoral VOIs (VOI_P) were automatically generated by expanding VOI_I by 1, 2, 3, and 4 mm. A total of 107 radiomic features were extracted from each VOI. Feature selection was conducted using the t test, Spearman correlation analysis, and the least absolute shrinkage and selection operator. Subsequently, radiomics models (VOI_I model, VOI_P model, and a combined model) were constructed. The models' performance was assessed by the area under the curve (AUC). RESULTS: In the training set, the intratumoral (VOI_PET_I) model achieved the highest AUC (0.973), followed by the peritumoral 1 mm (VOI_PET_P1) model (AUC = 0.952). However, in the independent testing set, the VOI_PET_P1 model outperformed the VOI_PET_I model, achieving the highest AUC of 0.931 versus 0.920, respectively. This suggests that the peritumoral 1 mm region may provide superior generalizability for subtype classification in unseen data. CONCLUSION: In this study, radiomics features extracted from both intratumoral and peritumoral regions on 18F-FDG PET/CT demonstrated high predictive value for classifying NSCLC pathological subtypes. Notably, the 1 mm peritumoral region exhibited the best performance for differentiating lung adenocarcinomas, especially in the independent testing cohort. However, these findings are based on a single-center cohort with a limited sample size and require further validation in larger, multicenter studies to confirm their clinical utility and generalizability.