Deep Learning-Based Real-Time Classification of Thoracic Pathologies in Chest Radiographs

Abstract

Diagnosing thoracic diseases from chest radiographs remains challenging, especially in resource-limited environments. This study presents YOLOv8n-cls, a lightweight deep learning model for real-time classification of five pathologies: 1) COVID-19, 2) fibrosis, 3) normal, 4) pneumonia, and 5) tuberculosis. The model was trained on a dataset of 11,019 chest X-ray images, combining public data from NIH ChestX-ray14 and a private clinical dataset, and achieving a Top-1 accuracy of 92.23%. Preprocessing included format conversion and text removal, while data augmentation techniques such as flipping, rotation, brightness/contrast adjustment, and affine translation were applied to improve model generalization. Performance evaluation relied on confusion matrices, precision, recall, F1-score, specificity, and ROC-AUC curves. Moreover, Grad-CAM visualizations were employed to enhance interpretability and analyze misclassification patterns. YOLOv8n-cls provides a strong balance between accuracy and computational efficiency, making it suitable for real-time clinical deployment.