Radiomics‐Driven Lung Adenocarcinoma Subtype Classification

ObjectiveThis study aimed to identify the optimal classification model for lung adenocarcinoma (LUAD) subtypes through radiomics‐driven analysis, addressing challenges such as data set imbalance, small sample sizes, and the need for accurate multi‐class classification.MethodsRadiomic features were extracted from CT scans and integrated with machine‐learning and deep‐learning techniques to improve diagnostic accuracy. After preliminary feature selection, the most effective feature subsets were identified by comparing single‐stage and multi‐stage feature selection methods, such as recursive feature elimination (RFE), random forest (RF), and Lasso. SMOTE techniques were applied to address class imbalance through data augmentation, and loss functions such as cross‐entropy were used for model training and evaluation. Finally, classification was performed using RF, KNN, GBDT, SVM, Stacking, Voting, and deep‐learning models (ResNet‐18, ResNet‐50, VGG16, etc.).ResultsThe MStacking model, based on mutual information (MI) and the stacking ensemble algorithm, achieved superior performance with a classification accuracy of 82.00%, precision of 82.00%, F1 score of 83.00%, AUC of 95.00%, sensitivity of 79.00%, and specificity of 94.00%. These results outperformed other methods. Deep‐learning models showed limited performance when trained on small sample sizes. However, when integrated with radiomics features, CNN models, particularly ResNet‐50, demonstrated significantly improved performance, especially when addressing class imbalance using SMOTE, with ResNet‐50's accuracy increasing by 20%. The MStacking model also showed stable performance in multi‐class tasks.ConclusionRadiomics‐driven deep‐learning models demonstrated a significant advantage in LUAD subtype classification, particularly when dealing with small sample sizes. Integrating radiomics features enhanced the performance of deep‐learning models, offering a promising approach for LUAD classification.