AI Tool Shows High Accuracy in TB Drug-Resistance Screening, ICMR-NIRT Study Finds

While rapid detection of multidrug-resistant (MDR) or rifampicin-resistant (RR) TB is vital for patient care, the National Tuberculosis Elimination Program (NTEP) currently relies on manual LPA interpretation for first-line (FL) and second-line (SL) drugs, which is often limited by subjective variability. Consequently, Sucharitha Kannappan Mohanvel and colleagues of the Indian Council of Medical Research-National Institute for Research in Tuberculosis (ICMR-NIRT) aimed to establish and verify an automated AI system to improve the precision, uniformity, and scalability of resistance reporting across diverse clinical settings.

Between 2023 and 2024, the ICMR-NIRT validated an AI system—integrating Faster Regions Convolutional Neural Network (FR-CNN), Detection Transformer (DETR), and Hierarchical Neural Network (HNN)—across three phases using 2,810 FL and 241 SL LPA strips from ten intermediate reference laboratories (IRLs). Only samples with valid internal quality control (IQC) and confirmed Mycobacterium tuberculosis (MTB) detection were analyzed to assess diagnostic accuracy and the F1 score.

Key Findings of the Study Include:

• Exceptional Diagnostic Accuracy: The study evaluation demonstrated high diagnostic precision with an accuracy range of 92 percent to 100 percent and an overall F1 score between 0.81 and 1.00 across all targeted resistance genes.
• High Predictive Reliability: The tool achieved nearly perfect negative predictive value (NPV) and positive predictive value (PPV) scores, ranging from 0.99 to 1.00 for the tub band and major drug-resistance genes like rpoB, katG, and inhA
• Robust Sensitivity and Specificity: Clinical performance remained strong across all validation phases, maintaining a sensitivity of 80 percent to 100 percent and a specificity of 82 percent to 100 percent for both FL and SL drugs.
• Minimal Expert Intervention: While human-in-the-loop review was initiated for roughly one-third of samples, expert microbiologists only needed to correct AI-generated interpretations in 3 to 9 percent of total cases, typically due to faint bands or artifacts.
• Consistent Performance Across Hardware: The study reveals that the AIsolution maintained reliable results across various scanner resolutions (260–600 dots per inch (dpi)), ensuring that variations in imaging equipment do not compromise the quality of drug resistance detection.

The results suggest that this AI system provides a reliable, expert-level alternative for TB drug resistance screening, achieving an accuracy range of 92–100 percent and a near-perfect F1 score reaching 1.00. By integrating this tool with expert oversight, laboratories can significantly enhance result uniformity and reduce interpretation time, directly supporting national goals for faster treatment delivery and disease elimination.

The authors that implementing this AI system streamlines laboratory workflows and ensures uniform LPA reporting, thereby accelerating the initiation of appropriate treatment regimens and supporting national TB elimination efforts.

Reference

Mohanvel SK, Radhakrishnan R, Balraj P, et al. Artificial intelligence for screening drug resistance in tuberculosis. Indian J Med Res 2025; 162: 910-916.