Integrating AI with PCR for Tuberculosis Diagnosis: Evaluating a Deep Learning Model for Chest X-Rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• Tuberculosis (TB) poses a significant global health threat, demanding efficient diagnostic tools.
• Chest radiography (CXR) is a common but non-specific screening method for TB.
• Polymerase chain reaction (PCR) assays are crucial for confirming TB diagnoses.

Purpose of the Study:

• To assess the diagnostic performance of a deep learning model (DLM) for TB detection via CXR.
• To compare the DLM's accuracy against PCR results in a low-burden region.
• To evaluate the DLM's utility in resource-limited settings.

Main Methods:

• A retrospective analysis of CXR images and PCR findings from two hospitals.
• Utilized a CheXzero vision transformer-based DLM trained on extensive imaging data.
• Evaluated performance using receiver operating characteristic (ROC) curves, area under the curve (AUC), sensitivity, and specificity.

Main Results:

• The DLM achieved high internal (AUC 0.915) and external (AUC 0.850) validation scores.
• The model demonstrated good sensitivity and specificity, though accuracy decreased for PCR-confirmed cases only.
• Performance was reduced in older adults and individuals with specific comorbidities.

Conclusions:

• AI-assisted CXR screening offers potential for TB detection, especially in resource-limited areas.
• The DLM shows promise as a supportive tool for TB screening.
• Confirmatory PCR testing remains indispensable for definitive TB diagnosis.