Abstract Chest radiog raphy is central to tuberculosis (TB) screening, yet automated triage remains challenging because pathological cues are subtle and datasets are imbalanced. This study presents a lightweight, explainable pipeline that fuses CNN semantic embeddings with Histogram of Oriented Gradients (HOG) texture features and classifies the joint vector using a leaf-wise LightGBM ensemble. Images are resized, denoised (BM3D), contrast-enhanced (CLAHE), and intensity-normalized; CNN and HOG features are concatenated and scored with histogram splits, exclusive feature bundling, gradient-based one-side sampling, and depth-capped growth. Probabilities are optionally calibrated on a validation fold, and Grad-CAM computed from the convolutional branch provides case-level visual explanations. Evaluation uses a curated public TB chest-X-ray subset (8000 images after augmentation) with a held-out test partition of original radiographs (n = 1400). On the test cohort, the proposed fusion attains 99.45% accuracy, 99.45% precision, 99.10% recall (sensitivity), 99.84% specificity, and a 99.35% F1-score, matching the tabulated results. Compared with single-source extractors (HOG-only and ConvNeXt-V2), the approach reduces both false positives and false negatives, exhibits rapid and stable convergence, and localizes clinically plausible regions. The findings indicate that combining semantic and gradient-texture information with an efficient, calibratable leaf-wise boosted classifier delivers state-of-the-art, interpretable TB screening on chest radiographs while remaining computationally practical.
Islam et al. (Sun,) studied this question.