A deep learning-based algorithm for pulmonary tuberculosis detection in chest radiography.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Deep Learning for Diagnostics

Background:

• Tuberculosis (TB) diagnosis using chest radiography (CXR) is challenging due to variable patterns mimicking other conditions.
• Accurate and efficient diagnostic tools are crucial for timely TB detection and management.

Purpose of the Study:

• To evaluate the efficacy of a deep neural network (DNN) algorithm developed with Google Teachable Machine for predicting TB probability from CXRs.
• To compare the AI algorithm's performance against human expert readers and radiological reports.

Main Methods:

• A DNN-based image classification tool was trained using a dataset of 348 TB CXRs and 3806 normal CXRs.
• External validation was performed on 250 CXRs, with performance compared to five pulmonologists.
• The algorithm's accuracy was assessed in detecting TB and differentiating it from other abnormal CXRs.

Main Results:

• The AI algorithm achieved high areas under the curve (AUC) of 0.951 and 0.975 in external validation datasets.
• AI performance was comparable to experienced pulmonologists (AUC 0.936-0.995).
• Combining AI with human readers improved diagnostic accuracy (AUC 0.862-0.885) compared to either alone.

Conclusions:

• The TB CXR AI algorithm developed using Google Teachable Machine is effective for TB detection.
• The AI shows accuracy close to experienced clinicians and can be a valuable tool for CXR-based TB diagnosis.