Intelligent diagnostic model for pulmonary disease based on auto-metric graph neural network with chest CT images

Non-small cell lung cancer (NSCLC) exhibits pronounced biological heterogeneity and intricate spatial architecture, posing significant challenges for image-based tumor subtyping and prognostic evaluation. In this study, we present a classification framework that integrates three-dimensional convolutional neural network (3D CNN) feature extraction with the Auto-Metric Graph Neural Network (AMGNN) to facilitate subtype identification in NSCLC patients. High-dimensional structural features are first extracted from chest CT scans, followed by dimension reduction using LASSO, PCA, and a hybrid approach combining both techniques. Patient similarity graphs are then constructed based on feature proximity, enabling end-to-end graph-based classification through AMGNN. Experiments conducted on the publicly available Lung1 dataset demonstrate that the AMGNN model consistently outperforms traditional classifiers such as support vector machines (SVM), random forests (RF), and gradient boosting (XGBoost) in terms of accuracy and robustness. These findings highlight the superior capability of AMGNN in modeling complex, high-dimensional medical imaging data, and underscore its potential in clinical decision support, particularly for fine-grained tumor subtyping in heterogeneous malignancies.