Development and Validation of Deep Learning-Based Infectivity Prediction in Pulmonary Tuberculosis Through Chest...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Infectious Disease Diagnostics

Background:

• Pulmonary tuberculosis (PTB) diagnosis is hindered by lengthy traditional methods like smear and culture tests.
• These conventional tests require hours to weeks for results, posing a global health challenge.
• Rapid and accurate diagnostic tools are crucial for effective PTB management and control.

Purpose of the Study:

• To develop and validate an AI-based system using chest radiography (CXR) for rapid PTB infectivity assessment.
• To compare the AI model's performance against traditional smear and culture tests for speed and accuracy.
• To leverage AI for improved efficiency in PTB screening and diagnosis.

Main Methods:

• Utilized DenseNet121 architecture with TransUNet for lung region segmentation and cropping on 36,142 CXR images from 4492 PTB patients.
• Employed visualization techniques (grad-CAM, LIME) to interpret AI model decision-making.
• Validated the model using internal (2022-2023) and external (Yongin Severance Hospital) datasets.

Main Results:

• Internal validation showed 73.27% accuracy, 0.79 ROC AUC, and 0.77 PR AUC.
• External validation achieved 70.29% accuracy, 0.77 ROC AUC, and 0.8 PR AUC.
• AI model's decision-making process was elucidated using visualization techniques.

Conclusions:

• The AI tool provides a swift and precise method for assessing PTB infectivity via CXR.
• This AI application can significantly improve screening efficiency by enabling pre-sputum test infectivity evaluation.
• The AI-driven approach offers a promising alternative to traditional diagnostic methods in clinical practice.