Machine learning-based early detection of tuberculosis in asymptomatic high-risk populations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Public Health

Background:

• Tuberculosis (TB) is a significant global health issue, with dormant cases fueling transmission.
• Current chest X-ray screening for TB faces challenges due to subtle radiographic signs and limited radiologist expertise.
• Deep learning models often require large datasets and struggle with global context in medical images.

Purpose of the Study:

• To develop and evaluate a machine learning framework for early TB detection in asymptomatic, high-risk populations.
• To improve the accuracy and interpretability of automated TB screening using chest X-rays.
• To address limitations of conventional deep learning models in capturing global image context.

Main Methods:

• Proposed a framework combining contrastive pretraining with a fine-tuned vision transformer (CPT-TB).
• Utilized a publicly available chest X-ray dataset (4200 images: 3500 normal, 700 TB).
• Employed self-attention mechanisms to aggregate information across the entire lung field for subtle pattern detection.

Main Results:

• CPT-TB achieved an AUC of 98.2%, accuracy of 95.5%, sensitivity of 94.9%, and specificity of 96.0% in five-fold cross-validation.
• Demonstrated a 4.0% accuracy increase over ResNet-50 and a 2.3% improvement over a supervised vision transformer.
• Attention maps provided interpretable visual cues for clinicians.

Conclusions:

• The CPT-TB framework shows significant promise for accurate and scalable TB screening.
• This approach can facilitate active case finding, especially in resource-limited settings.
• The model's interpretability aids clinical decision-making in TB diagnosis.