Explainability-Driven Tuberculosis Detection: GAN-Augmented CNN–Swin Transformer Hybrid Framework with Signal Processing Insights

Globally, tuberculosis (TB) remains a significant health concern, especially in areas with inadequate resources where prompt and precise diagnosis is difficult. The article presents a novel hybrid deep learning architecture that combines Generative Adversarial Networks (GANs), Convolutional Neural Networks (CNNs), and Swin Transformers to answer the demand for reliable and interpretable automated TB identification. By enhancing the quality of the input data, extracting both local and global features, and producing visually understandable outputs to aid in clinical decision-making, the architecture improves diagnostic accuracy.There are three stages in the proposed pipeline. In order to increase training stability and the size of the dataset, a DCGAN first creates high-quality synthetic chest X-ray (CXR) pictures. Second, lung areas are segmented by an attention-guided U-Net, which isolates pertinent anatomical features. Lastly, a CNN branch (fine-grained spatial features) and a Swin Transformer branch (contextual and long-range dependencies) are combined in a dual-branch classification module. Grad-CAM and attention heatmaps draw attention to discriminative lung regions that affect forecasts in order to guarantee interpretability.In-depth analyses were performed on three benchmark datasets: Montgomery, Shenzhen, and TBX11K. The suggested model outperformed traditional CNN- or Transformer-only baselines by achieving 99.2% accuracy, 0.985 F1-score, and 0.98 AUC on the TBX11K dataset. The Shenzhen and Montgomery datasets showed comparable gains, demonstrating the generalizability and resilience of the methodology. In addition to improving classification performance, the combination of CNN–Transformer fusion and GAN-based data improvement makes clinical adoption easier by producing explainable results. A promising step toward transparent, scalable, and trustworthy AI-assisted TB diagnosis is this hybrid approach.