Deep learning-based pulmonary tuberculosis automated detection on chest radiography: large-scale independent testing

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• External validation of deep learning systems is crucial for clinical prognoses.
• Assessing artificial intelligence (AI) performance in detecting tuberculosis (TB) on external datasets is essential.

Purpose of the Study:

• To evaluate the performance of an AI system for tuberculosis detection.
• To validate the AI system on a large-scale, independent external dataset.

Main Methods:

• A deep convolutional neural network (DCNN) was developed using an internal dataset of 7,025 chest X-ray images.
• The AI system underwent external validation on a 6-year cohort dataset comprising 358,169 images.

Main Results:

• The AI system achieved an AUC of 0.95 and accuracy of 0.931 for TB detection on the external dataset.
• The system demonstrated high performance in delineating lung regions (Dice coefficient 0.958) and classifying images.

Conclusions:

• The AI system is a reliable tool for real-world independent validation.
• The validated AI system can serve as an effective TB screening tool, aiding radiologists in flagging potential cases.