Multi-view deep learning framework for the detection of chest X-rays compatible with pediatric pulmonary tuberculosis

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pediatric Infectious Diseases
• Radiology and Imaging Analysis

Background:

• Tuberculosis (TB) is a significant global health issue, with pediatric TB diagnosis being particularly challenging due to non-specific symptoms and subtle radiological findings.
• Many children with TB remain undiagnosed or untreated, highlighting the need for improved diagnostic tools.
• The World Health Organization (WHO) recommends chest X-rays (CXRs) for TB screening due to their accessibility and rapid assessment capabilities.

Purpose of the Study:

• To develop and evaluate pTBLightNet, a multi-view deep learning framework for detecting pediatric pulmonary TB from CXRs.
• To assess the framework's performance across different age groups and its generalizability using various reference standards.
• To explore the impact of incorporating lateral CXR views on diagnostic accuracy in children.

Main Methods:

• Developed pTBLightNet, a multi-view deep learning framework utilizing both frontal and lateral CXR views.
• Pre-trained the model on a large adult CXR dataset (N=114,173) and fine-tuned/trained it on pediatric TB cohorts (N=918).
• Evaluated the framework's performance using internal and external testing, comparing against CXR TB compatibility, expert readings, microbiological confirmation, and case definitions.

Main Results:

• Achieved an area under the curve (AUC) of 0.903 on internal testing and 0.682 on external testing.
• External evaluation demonstrated effectiveness and generalizability across diverse reference standards.
• Age-specific models (<5 and 5-18 years) performed competitively, with lateral CXRs improving diagnosis in younger children.

Conclusions:

• pTBLightNet demonstrates robustness and potential for improving pediatric TB diagnosis, particularly in resource-limited settings.
• The framework's ability to leverage multi-view CXRs and age-specific models enhances its diagnostic utility.
• This deep learning approach offers a promising avenue for addressing the challenges in pediatric TB detection.