Drug-Resistant Tuberculosis Treatment Recommendation, and Multi-Class Tuberculosis Detection and Classification Using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Tuberculosis (TB) diagnosis and drug resistance classification remain critical global health challenges.
• Current diagnostic methods can be time-consuming and may lack precision in identifying drug-resistant strains.
• The need for rapid, accurate, and accessible diagnostic tools for TB and its resistant forms is paramount.

Purpose of the Study:

• To develop a decision support system (TB-DRC-DSS) for detecting tuberculosis and categorizing drug-resistant strains using deep learning.
• To create an ensemble model integrating multiple Convolutional Neural Network (CNN) architectures for enhanced diagnostic accuracy.
• To provide a web-based platform for real-time TB and drug resistance classification.

Main Methods:

• Development of a deep learning ensemble model using EfficientNetB7, MobileNetV2, and Dense-Net121 architectures.
• Application of image segmentation, data augmentation, and decision fusion techniques to optimize classification performance.
• Training and validation on a combined dataset of 7,008 chest X-ray images from multiple public sources.
• Implementation of a web application for user interaction and diagnosis.

Main Results:

• The TB-DRC-DSS demonstrated significant improvements in classifying drug-sensitive TB (DS-TB) against drug-resistant TB (DR-TB) by an average of 43.3%.
• Enhanced accuracy in differentiating between DS-TB and MDR-TB (28.1%), DS-TB and XDR-TB (6.2%), and MDR-TB and XDR-TB (9.4%).
• The multiclass model achieved a high accuracy of 92.6% on the test dataset and 92.8% on a random subset, with a user preference score of 9.52/10 from medical staff.

Conclusions:

• The proposed deep learning ensemble model offers a highly accurate and efficient solution for TB and DR-TB detection and classification.
• The web-based TB-DRC-DSS provides a valuable tool for clinicians, improving diagnostic capabilities and potentially patient outcomes.
• The system's performance and positive user feedback indicate its potential for widespread clinical adoption.