Combining semantic CNN and HOG texture with LightGBM for explainable tuberculosis detection

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.