Integrating Multi-Source Pulmonary X-Ray Data into a Cross-Validated CNN with AdamW and Augmentation
TL;DR
This study presents an approach for classifying lung diseases using a tailored convolutional neural network (CNN) that combines several chest X-ray (CXR) datasets and applies extensive data augmentation to build a balanced dataset and improve generalization.
Abstract
The global spread of COVID-19 has affected health and economic conditions worldwide, and new variants continue to appear despite the availability of vaccines. These variants share a common effect on the respiratory system, which keeps pulmonary disease detection a major priority. This study presents an approach for classifying lung diseases using a tailored convolutional neural network (CNN). The system combines several chest X-ray (CXR) datasets and applies extensive data augmentation to build a balanced dataset and improve generalization. The model distinguishes five classes: Normal, COVID-19, Tuberculosis, Viral Pneumonia, and Bacterial Pneumonia. Cross-validation ensures more stable performance and reduces overfitting risks. The method uses the AdamW optimizer to improve convergence and training stability. Performance evaluation includes confusion matrix indicators such as accuracy, specificity, precision, sensitivity, and F1-score, along with ROC curves. The results show that the method is reliable and efficient for automatic classification of CXR images. The CNN model had an average accuracy of 95.26% and a macro-AUC of 99.32% across five folds. Our model is implemented on the TensorFlow framework with the datasets that are public and available for the research community.