A multi-stage deep learning approach for comprehensive lung disease classification from x-ray images

Lung diseases are among the leading causes of morbidity and mortality worldwide, imposing a significant burden on healthcare systems and affecting millions of individuals annually. Despite the rapid advancements in artificial intelligence, current studies on automated lung disease diagnosis often focus on a limited subset of diseases, typically 3–5 classes, which restricts their applicability in real-world scenarios. Furthermore, these studies frequently overlook the role of hyperparameter optimization in enhancing the predictive accuracy of deep learning models. In this study, we propose a multi-stage CNN toolchain capable of classifying 17 distinct lung conditions. Each stage refines the classification by identifying subsets of diseases based on their etiology and co-occurrence tendencies, ultimately narrowing down to specific diagnoses. Five pre-trained CNN architectures, including ResNet-50, DarkNet-53, EfficientNet-b0, ResNet-101, and DenseNet-201, are systematically compared. Hyperparameters such as solver type, batch size, and learning rate are optimized to maximize performance, measured using standard metrics such as accuracy, precision, recall, F1-score and specificity. The results are benchmarked against a recent study, demonstrating the efficacy of our approach in tackling a wide-ranging multi-class classification problem. This work underscores the importance of a multi-stage approach and hyperparameter tuning in developing robust, scalable diagnostic tools for lung disease detection. The proposed method not only improves diagnostic accuracy but also provides a foundation for future research on generalizing AI models for comprehensive lung disease classification.