Vit-Ensemble: Probabilistic voting based ensemble of Vision Transformers for tuberculosis detection using radiographs

Tuberculosis (TB) detection from chest X-ray (CXR) images remains a critical challenge in global healthcare. This study introduces Vit-Ensemble, a novel ensemble model leveraging Vision Transformer (ViT) architectures for robust TB detection. The core innovation of Vit-Ensemble lies in its probabilistic voting strategy. Instead of relying solely on predicted class labels, it combines the probabilistic outputs of multiple ViT models trained on diverse TB datasets. By averaging these class probabilities, Vit-Ensemble makes decisions based on the collective confidence of individual models, enhancing generalization performance and reducing the impact of biases and uncertainties. To further improve diagnostic results, the authors systematically explore various image preprocessing techniques, including contrast enhancement and noise reduction. Through rigorous experimentation on benchmark datasets, Vit-Ensemble demonstrates superior performance compared to state-of-the-art convolutional neural network models. It achieves a remarkable 99.67% accuracy, outperforming both its individual components (e.g., DeiT-Base at 99.14%) and established convolutional neural networks (i.e., EfficientNet-B3 (99.64%) and DenseNet201 (93.21%)). Our results highlight the effectiveness of probabilistic voting within the ensemble framework for TB detection, offering the potential for early diagnosis and effective disease management. This research provides valuable insights into the integration of ViT architectures, probabilistic voting-based ensemble learning for medical image analysis. Vit-Ensemble represents a significant advancement in computer-aided TB diagnosis, promising improved public health outcomes and streamlined TB control efforts.