Quantitative chest computed tomography-based hybrid deep learning for chronic lung disease detection and staging

Abstract

This research proposes a hybrid deep learning model to accurately detect chronic lung diseases in chest computed tomography (CT) scans. Current deep learning methods, such as convolutional neural networks (CNNs) and transformer models, often suffer from high computational requirements and low interpretability. This research combines a pre-trained residual network 50 layers (ResNet50) for deep feature extraction with hand-crafted texture features. The features are combined via hybrid concatenation, and the model is classified using an optimized extreme gradient boosting (XGBoost) ensemble with hyperparameter tuning to improve performance and reduce overfitting. This research compared the proposed model with radiomics-only, ResNet50 CNN, vision transformer (ViT-B16), and Swin transformer methods. This research used a public dataset containing adenocarcinoma, squamous cell carcinoma, large cell carcinoma, and normal classes. The classification accuracy of the hybrid framework was 92.38%, with better precision, sensitivity, specificity, F1-score, and receiver operating characteristic-area under the curve (ROC-AUC) than the baselines. A dual-mode explainable artificial intelligence (XAI) framework integrating gradient-weighted class activation mapping (Grad-CAM) and Shapley additive explanation (SHAP) was implemented to enable spatial lesion localization and feature attribution for clinical transparency. Grad-CAM was able to localize the lesions successfully, and SHAP revealed the influence of entropy-based texture features and semantic embeddings on the predictions. Overall, the framework develops advanced XAI systems for chest CT analysis.