Artificial intelligence for screening drug resistance in tuberculosis.

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Healthcare
• Tuberculosis Research

Background:

• The Central TB division developed an artificial intelligence (AI) tool to interpret line probe assay (LPA) strips.
• Over 18,000 LPA strips were used for training and validation across culture and drug susceptibility testing laboratories.
• Independent evaluation by ICMR-NIRT was conducted to verify the AI tool's performance.

Purpose of the Study:

• To establish and validate an AI-driven system for automated interpretation of LPA strips.
• To enhance the accuracy, consistency, and scalability of tuberculosis drug resistance screening.
• To support diverse laboratory settings within India's TB programme.

Main Methods:

• The AI system utilizes Faster R-CNN for strip detection, DETR for band localization, and a hierarchical neural network (HNN) for classification.
• Independent validation involved 2810 first-line (FL-LPA) and 241 reflex second-line (SL-LPA) tests across ten intermediate reference laboratories.
• Performance was assessed using metrics including accuracy, sensitivity, specificity, and F1 score for key genes.

Main Results:

• AI models achieved accuracy ranging from 92-100%, sensitivity from 80-100%, and specificity from 86-100%.
• The overall F1 score ranged from 0.81 to 1.00, demonstrating high precision and recall.
• The AI tool demonstrated expert-level interpretation capabilities for LPA strips.

Conclusions:

• The AI system provides a novel, modular, and scalable solution for LPA interpretation, performing on par with expert readers.
• Adoption of this AI tool can significantly reduce interpretation time and enhance result uniformity.
• The AI tool supports national goals for TB elimination by improving treatment delivery across India.