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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Tuberculosis (TB) detection from chest X-ray (CXR) images is crucial for global health.
• Existing methods face challenges in accuracy and generalization.

Purpose of the Study:

• To develop a robust and accurate model for TB detection from CXR images.
• To leverage Vision Transformer (ViT) architectures and ensemble learning for improved diagnostic performance.

Main Methods:

• Developed Vit-Ensemble, an ensemble model using multiple ViT architectures.
• Implemented a probabilistic voting strategy by averaging class probabilities from individual ViT models.
• Explored various image preprocessing techniques, including contrast enhancement and noise reduction.

Main Results:

• Vit-Ensemble achieved a high accuracy of 99.67% on benchmark datasets.
• The model outperformed individual ViT components (DeiT-Base: 99.14%) and state-of-the-art CNNs (EfficientNet-B3: 99.64%, DenseNet201: 93.21%).
• Probabilistic voting enhanced generalization and reduced model bias.

Conclusions:

• Vit-Ensemble demonstrates superior performance for TB detection compared to existing methods.
• The study highlights the effectiveness of probabilistic voting in ensemble learning for medical image analysis.
• This advancement offers potential for earlier TB diagnosis and improved disease management.