Assessing Invasiveness of Ground-Glass Nodules Using Ternary-Class CT Radiomics Models: A Multi-Center Study with SHAP Explanations.

Background: Ground-glass nodules (GGNs) exhibiting varying degrees of invasiveness necessitate distinct clinical management protocols and therapeutic interventions. This study aimed to construct ternary classification machine-learning models utilizing CT radiomics features to stratify GGNs into precursor glandular lesions (PGL), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC). Methods: This multi-center and retrospective study analyzed 1130 GGNs patients, comprising 858 cases (mean age: 57.635±12.978 years) in the training set and 272 cases (mean age: 57.037±11.683 years) in the testing set. Computed tomography (CT) based radiomics features were extracted and utilized for machine-learning model construction. Six ternary classification models were developed, including Logistic Regression, RandomForest, ExtraTrees, XGBoost, LightGBM, and multi-layer perceptron (MLP) models, to establish a comprehensive multi-class prediction framework. The area under the receiver operating characteristic curve (AUC) serves as a quantitative metric for evaluating model performance. Results: The MLP-Model demonstrated superior predictive performance, with ternary classification accuracy reaching 0.712 in the training set and maintaining 0.658 in the testing set, surpassing the performance range of comparative models (training set: 0.590-0.690; testing set: 0.577-0.614). The MLP-Model achieved micro and macro AUCs of 0.877 (95% CI: 0.863-0.890) and 0.861 (95% CI: 0.835-0.886) in the training set, and 0.808 (95% CI: 0.776-0.840) and 0.799 (95% CI: 0.740-0.854) in the testing set. Conclusion: Our study developed ternary machine-learning models, particularly the MLP-Model, effectively stratifies GGN invasiveness into PGL, MIA, and IAC subtypes, thereby optimizing clinical decision-making through precision therapeutic planning and personalized management strategies.