Peritumoral features for assessing invasiveness of lung adenocarcinoma manifesting as ground-glass nodules

Abstract

We evaluated the predictive value of radiomics features from different peritumoral ranges for the invasiveness of ground-glass nodular lung adenocarcinoma using various machine learning models. This retrospective study included 317 patients with 323 ground-glass nodules diagnosed as minimally invasive adenocarcinoma (MIA) or invasive adenocarcinoma (IAC) at Benxi Central Hospital (January 2019–December 2023). Radiomic features from tumor margins of 1, 2, 3, 4, and 5 mm were extracted. Eight machine learning models were constructed following dimensionality reduction. The models were evaluated using receiver operating characteristic curves. All models had area under the curve values > 0.75, effectively distinguishing between MIA and IAC. Only the decision tree model showed statistically significant differences (P < 0.05); no differences were found between the other models (P > 0.05). In the training set, the 1-mm margin model achieved the highest ranking, followed by the 2-, 4-, 5-, and 3-mm models. In the validation group, the 3-mm margin model ranked the highest, followed by the 2-, 1-, 4-, and 5-mm models, with no statistically significant differences (P > 0.05). All machine learning models demonstrated good predictive performance for both MIA and IAC. Radiomic features from 1 to 5-mm margins showed strong predictive value, though no optimal margin range was identified.