An efficient channel attention–based lightweight network for tuberculosis screening from chest X-ray images

Abstract Tuberculosis remains a major global health burden, and timely screening with chest X-ray imaging is particularly important in resource-limited clinical settings. However, many deep learning approaches for automated tuberculosis detection are constrained by class imbalance, computational overhead, and limited interpretability. This study proposes LiteTBNet-ECA, a lightweight and interpretable framework for binary classification of tuberculosis versus normal chest X-ray images. The model adopts MobileNet-style inverted residual design with depthwise separable convolutions for computational efficiency and incorporates Efficient Channel Attention to strengthen discriminative feature representation with minimal added cost. An end-to-end workflow is established, covering standardized preprocessing, imbalance-aware training, evaluation on a held-out test set, and post hoc interpretability. Images are resized to 224 × 224 and normalized; augmentation is applied only to the training subset, and strict separation of training, validation, and testing partitions is maintained to mitigate data leakage. Performance is examined under five imbalance-handling settings: no oversampling, weighted averaging, SMOTE, ADASYN, and Borderline-SMOTE. Across these settings, LiteTBNet-ECA demonstrates robust performance, achieving under ADASYN a test accuracy of 99.56%, precision 99.48%, recall 99.62%, F1-score 99.55%, and ROC-AUC 1.00. Grad-CAM visualizations highlight lung regions contributing to predictions, supporting qualitative interpretability without implying lesion-level delineation. Overall, LiteTBNet-ECA provides an accurate and efficient tuberculosis screening approach with strong interpretability characteristics, supporting its potential use in screening-oriented workflows.