Tuberculosis Diagnostics and Localization in Chest X-Rays via Deep Learning Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Tuberculosis (TB) remains a significant global health threat, causing numerous deaths annually.
• Chest X-rays (CXRs) are the primary, cost-effective tool for TB detection, but diagnostic accuracy can be improved.
• Automated analysis of CXRs can enhance TB detection and localization of affected lung areas.

Purpose of the Study:

• To develop an efficient computer-aided detection (CAD) system for TB diagnosis using CXRs.
• To improve TB diagnostic accuracy by integrating deep learning models for detection and localization.
• To support radiologists and clinicians in making informed TB diagnoses from CXRs.

Main Methods:

• Proposed an integrated process using convolutional neural networks (CNNs) and deep learning for TB detection and localization in CXRs.
• Implemented three key steps: CNN model structure modification, artificial bee colony algorithm for fine-tuning, and a linear average-based ensemble method.
• Evaluated performance on Shenzhen Hospital and NIH CXR datasets, validating detection of abnormalities and seven TB-related manifestations.

Main Results:

• The developed system demonstrated outstanding performance in detecting lung abnormalities on CXRs.
• Accurately differentiated among seven specific TB-related manifestations, including consolidation, effusion, fibrosis, infiltration, mass, nodule, and pleural thickening.
• Utilized class activation mapping for visual interpretation and localization of detected lung abnormalities on CXRs, surpassing state-of-the-art methods.

Conclusions:

• The proposed deep learning approach significantly enhances TB diagnostics through improved detection and localization on CXRs.
• The system offers a valuable tool for clinicians, improving the accuracy and efficiency of TB diagnosis.
• This research advances automated medical imaging analysis for infectious lung diseases like TB.