<![CDATA[Tuberculosis (TB) and pneumonia (PNA) are infectious lung diseases with overlapping chest X-ray (CXR) manifestations, making automated differential classification clinically important and methodologically challenging. This study proposes a supervised CXR classification workflow to distinguish TB from PNA using pretrained convolutional neural network (CNN) feature extractors and supervised classifiers. A publicly available de-identified dataset comprising 390 TB and 390 PNA images was used. Images were screened to exclude duplicates, corrupted files, non-CXR images, unclear labels, and identifiable cases. Preprocessing included format standardization, resizing according to CNN input requirements, and normalization. To reduce augmentation-based leakage risk, no heavy pre-validation augmentation was applied. Image embeddings were extracted using VGG-16, Inception V3, and VGG-19, then classified using Logistic Regression, Support Vector Machine, and Neural Network models. Performance was evaluated using stratified 5-fold cross-validation with AUC, accuracy, F1-score, precision, recall, MCC, and confusion matrix analysis. The Inception V3–Logistic Regression combination achieved the best performance, with AUC of 0.999, accuracy of 0.992, F1-score of 0.992, and MCC of 0.985.]]>
Copyrights © 2026
