Automated Pulmonary Tuberculosis Severity Assessment on Chest X-rays

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Tuberculosis (TB) remains a significant global health challenge, with millions falling ill and dying annually.
• The 2022 World Health Organization report indicated a concerning rise in TB cases and drug-resistant TB.
• Accurate assessment of pulmonary TB severity from chest X-rays (CXRs) is crucial for resource allocation and treatment monitoring, especially in resource-limited settings.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting the Timika score, a measure of TB severity.
• To compare three distinct deep learning approaches with varying levels of explainability for TB severity prediction.
• To assess the generalization performance of the best-performing model on unseen data.

Main Methods:

• Three deep learning approaches were proposed: 1) lesion detection (YOLOV5n) and cavitation prediction (DenseNet121) for score calculation. 2) Direct prediction of affected lung percentage (DenseNet121 regression) and cavitation presence (DenseNet121 classification). 3) Direct Timika score prediction (DenseNet121 regression).
• Models were trained and evaluated using chest X-ray data.
• Performance was assessed using mean absolute error and Pearson correlation on held-out datasets.

Main Results:

• The second approach, combining direct lung percentage and cavitation prediction, achieved the best performance.
• This approach yielded a mean absolute error of 13-14% and a Pearson correlation of 0.7-0.84.
• The models demonstrated good generalization capabilities on independent datasets.

Conclusions:

• Deep learning models can effectively predict pulmonary TB severity using chest X-rays.
• The proposed methods offer a promising tool for objective and efficient TB severity assessment.
• These AI-driven tools can support clinical decision-making and improve patient management in TB care.