Deep learning-based radiomics model to evaluate the invasiveness of pure ground-glass nodules

Objectives: The objective of this study is to develop a deep learning-based radiomics model for the assessment of the invasiveness of pure ground-glass nodules (pGGNs). Methods: A total of 1,517 patients with pathologically confirmed atypical adenomatous hyperplasia, adenocarcinoma in situ, minimally invasive adenocarcinoma, or invasive adenocarcinoma (IAC) were retrospectively enrolled from an initial cohort of 1,840 cases. Patients were divided into two groups: non-IAC (atypical adenomatous hyperplasia, adenocarcinoma in situ, and minimally invasive adenocarcinoma) and IAC. Using stratified sampling, the dataset was split into training, validation, and test cohorts in a 7:2:1 ratio. Three models, that is, clinical, deep learning (DeepL), and combined, were constructed based on the ResNet-101 architecture. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis, and statistical differences between models were assessed using the DeLong test. Results: In the test cohort, the AUCs of the clinical, DeepL, and combined models were 0.829 (95% confidence interval [CI]: 0.761–0.897), 0.864 (95% CI: 0.807–0.921), and 0.890 (95% CI: 0.839–0.942), respectively. The DeLong test identified statistically significant differences in AUC among the three models ( p < 0.05). Both the DeepL and combined models outperformed the clinical model in predicting the invasiveness of pGGNs. Conclusion: The combined model offers a non-invasive and efficient tool for evaluating the invasiveness of pGGNs, with the potential to support clinical decision-making.