Chest Radiography of Tuberculosis: Determination of Activity Using Deep Learning Algorithm.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Differentiating active tuberculosis (TB) from inactive, healed TB on chest radiographs is crucial in high TB incidence countries to prevent unnecessary treatment.
• Chest radiography (CR) is a common diagnostic tool, but distinguishing active from inactive TB requires expertise.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for estimating TB activity from a single chest radiograph.
• To assess the diagnostic performance of the DL model compared to human radiologists.

Main Methods:

• A convolutional neural network was fine-tuned using 3,824 active TB and 2,277 inactive TB CRs.
• The model was pretrained on pneumonia and normal chest X-ray datasets for feature learning.
• Performance was validated internally and on three external datasets, with ROC analysis comparing DL to radiologists.

Main Results:

• The DL model achieved high diagnostic performance with AUC values of 0.980 (internal) and up to 0.887 (external validation).
• On external validation, the DL model's AUC (0.815) was comparable to a thoracic radiologist (0.871) and superior to a general radiologist (0.811).

Conclusions:

• The developed DL algorithm demonstrates significant potential as an effective tool for identifying TB activity on chest radiographs.
• This AI-driven approach could improve patient follow-up and resource allocation in high TB burden regions.