Artificial intelligence for screening drug resistance in tuberculosis

AI interpretation mechanism

A classical computer vision technique (Fig. 1) is applied to identify and extract the strips pasted on the LPA sheet. Our AI pipeline, adapted from automatic interpretation of LPA for tuberculosis (2024) ¹¹, integrates Faster R-CNN for strip detection, DETR for band localisation, and an HNN for band classification. A transformer-based OCR (TrOCR) module is employed to extract patient or sample identifiers from each sheet. Once strips are extracted, as per the deep learning components, a bounding box (BB) is drawn around the band if a band is present in the strip. If two types of objects (foreground and background) are detected, models such as FR-CNN and DETR are applied to customise the architecture. Further, identifying the BB where bands are present, band types such as AC, CC, rpoB-Mut1, etc., are associated with each detected band and its corresponding band type. The BBs estimated by the DETR are sorted, and the relative distances of the centroids of the BBs from a reference are computed. Neural networks were trained using the Adam optimiser. Hyperparameters (such as learning rate, sequence length for long short-term memory (LSTM), batch size, and network depth) were optimised using grid search on the validation split. This forms the positional context for the band prediction model, where a sequence-to-sequence translation is carried out to create a bidirectional LSTM layer. Once band prediction is completed, the band probabilities are sent to the HNN and deterministic rules (DET) to produce drug interpretations. Band-level annotations (bounding boxes, presence/absence) were performed using a custom tool, then validated by multiple microbiologists. Final drug calls were derived either by consensus or deterministic rules applied to the band pattern. Some of the strip results are rejected based on HNN by the rejection classifier and DET and HNN by the discrepancy module. The ambiguous bands are highlighted by the fidelity module (FM) and sent for review by a human in the loop (trained microbiologist) for concordance ¹¹.

Close

Fig. 1.

Tuberculosis (TB) Line Probe Assay (LPA) Artificial Intelligence (AI) solution uses a deep learning-based computer vision model to automate LPA test interpretation. A Convolutional Neural Network (CNN) detects and classifies bands after preprocessing steps, including image enhancement, alignment correction, and segmentation. The model analyses band presence, position, and intensity to map patterns to wild-type and mutant regions, determining drug resistance. Trained on expert-annotated images, it achieves high accuracy, enabling automated, real-time LPA interpretation within diagnostic workflows.

Export to PPT

Study setting

The validation process by ICMR-NIRT was planned in three different phases:

1. Phase 1. The LPA sheets at ICMR -NIRT were scanned and interpreted by AI solution.

2. Phase 2. The LPA sheets from nine Intermediate reference laboratories (IRLs) (at Kolkata, Nagpur, Patna, Raipur, Telangana, Lucknow, Bengaluru, Chennai, and Jodhpur) were scanned using different types of scanners and sent to ICMR-NIRT for AI interpretation.

3. Phase 3. The LPA sheets that were subjected to scanning and AI interpretation at five IRLs (IRL Lucknow, Hyderabad, Bhopal, Chennai, and Nagpur) were cross-verified by ICMR-NIRT.

AI interpretation and validation

The AI validation conducted by Central TB division initially included 20,000 strips for training the microbiologists at 40 laboratories to perform testing and interpretation independently. Based on the consistency of results obtained, a minimum of 10 per cent of the trained set were included for final evaluation (2000 strips). Based on this calculation and accounting for a 20 per cent contamination rate, the minimum sample size required for this study was 2400 for FL-LPA results and their respective SL-LPA. In phase I, 1945 FL-LPA and 174 SL-LPA results were interpreted and analysed. In Phase II, 865 FL-LPA and 67 SL-LPA, and in Phase III, 271 FL-LPA and 75 SL-LPA, results were included for interpretation and analysis (Fig. 2). The analysis included all the LPA strips with satisfactory internal quality control (IQC) and MTB detection. Sheets with incomplete results or reporting blank/missing interpretations on LPA test results were excluded from the analysis. The minimum scanner specification required for the AI solution is 72 dpi and above; the scanners used in the IRLs across have a range of 260–600 dpi. Three levels of data prediction were used during the analysis; (a) the presence or absence of a band on a given strip; (b) interpretation of gene locus level; (c) the drug level interpretations.

Close

Fig. 2.

Flow of three phases of Artificial Intelligence First Line (FL)- and Second Line (SL)-Line Probe Assay (LPA) interpretation compared to the gold standard manual reading. Phase I-LPA strips were interpreted and analysed at the Indian Council of Medical Research-National Institute for Research in Tuberculosis (ICMR-NIRT), Phase II-Intermediate Reference Laboratories (IRLs) LPA strips at ICMR-NIRT and, Phase III- at ICMR-NIRT and the five IRLs.

Export to PPT

Discordance resolution

Microbiologists finalised the discordance resolution by comparing the LPA sheets with the field (nucleic acid amplification test) NAAT results, which is the usual diagnostic flow in FL-LPA interpretation. Such a manual interface is required for AI interpretation of false positives and negative results. The tubband representing MTB positive is considered regardless of its density or faintness. The LPA sheet, consisting of NAAT results obtained from the field, was verified by microbiologists during AI interpretation in the case of false-positive results. Lighter wild (WT) bands and mutant (MUT) bands were verified for false negatives and false positives, respectively, during the manual interface.

Statistical analysis

To assess the effectiveness of the AI model, F1 score^13-15, Cohen’s kappa, Fowlkes-Mallows score ¹⁶ and Matthew’s correlation coefficient (MCC) ¹⁷ were calculated and interpreted against the gold standard. By applying Cohen's Kappa, the agreement between the model's predictions and the gold standard was assessed. F1 Score was chosen to balance precision and recall, critical in minimising both false positives and false negatives in clinical diagnostics. MCC provides a robust, balanced evaluation in imbalanced datasets by incorporating all elements of the confusion matrix. Moreover, area under receiver operating curve (AUROC) reflects overall discriminative power, while area under plotted precision recall curve (AUPRC) is more informative when resistant cases (positive class) are rare, making it particularly relevant for gene-level resistance evaluation. To examine the dependability and resilience of the AI algorithm, the data obtained was analysed using the confusion matrix, ROC, and PRC ¹⁸. Respective sensitivity (recall), specificity, accuracy, negative predictive value, positive predictive value (precision), and F1 score, along with the respective AUROC and AUPRCwere obtained. The sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) were calculated to evaluate the performance of the AI model ¹⁹.

ICMR-NIRT evaluation

FL-LPA (n=1945) and SL-LPA (n=174) were extracted from the LPA sheets from the ICMR-NIRT routine diagnostics. Nine IRLs were included in the study, with a sample size of FL-LPA (n=860) and SL-LPA (n=67), respectively. Out of 174 cases from ICMR-NIRT, SL-LPA reported eight invalids and one indeterminate, while in 1945, FL-LPA strips had three reported invalids

Phase I

The NPV score for the tubband, gyrA/gyrB and rrs genes was 1.00; rpoB, KatG, and InhA were 0.99. The PPV scores for the tubband, rpoB, KatG, InhA, and gyrA/gyrB genes were 0.99-1.00. The accuracy, sensitivity, and specificity were compared for all performances of the tubband and the genes rpoB, katG, InhA, gyrA/gyrB genes, rrs, and eis. The accuracy ranges were as follows: overall (0.92-1.00), sensitivity (0.88-1.00), and specificity (0.86-1.00).The overall F1 score range from 0.89 to 1.00, showcasing ideal precision and recall for all the AI comparative model’s performance at ICMR-NIRT (Table). A scanner with a resolution of 600 dpi was used at ICMR-NIRT to scan the LPA sheets.

Phase II

For the Nine IRLs, specified alone,the NPV and PPV scores range between 0.99 and 1.00. The accuracy varied between 0.99 and 1.00, the sensitivity between 0.80 and 1.00, and the specificity between 0.98 and 1.00 for the tubband and the respective genes. The IRL scanners across had a 260–600 dpi resolution. The F1 score was similar to Phase I, even exhibiting the perfect PR for the AI tool compared with the gold standard.

Phase III

The performance of the AI and manual results at five IRL: NPV ranged from 0.95 to 1.00, and the PPV was 0.78–1.00 for the tubband and the rpoB, KatG, InhA, gyr A, gyrB, and rrs genes. The IRLs demonstrated performance levels ranging from 0.88 to 1.00, with sensitivity from 0.81 to 1.00 and specificity of 0.82 and 1.00 for the corresponding tubband and genes. The overall F1 score ranged from 0.81 to 1.00, showing ideal precision and recall. According to Kuang et al ¹⁹ (2022), a high F1 score (close to 1) indicated a balanced trade-off between precision and recall ²⁰ (Supplementary Figure).

Considering the minimum scanner specification required for the solution is 72 dpi and above, the scanners used in the five IRLs (Lucknow, Hyderabad, Bhopal, Chennai, and Nagpur) across had a range of 260–600 dpi (Supplementary Table). Based on the above inference, the scanning performed at IRLs and NRLs did not show any difference in the band density of LPA sheets, though there was a difference in the dpi across.

Despite timely outcomes, LPA data interpretation deficiencies persist in many heavily burdened laboratories. Integrating AI-driven systems enhances laboratory workflows for LPA testing, efficiently transmitting data to designated information systems for swift public health interventions. We aim to implement AI-LPA at C&DST laboratories nationwide to achieve the quickest treatment regimen. By utilising an AI solution, this leading study on LPA implementation in India aims to shorten interpretation time and increase the uniformity of the interpretation process across states. The current use of AI and advancements in ubiquitous computing for forecasting respiratory illnesses has opened up a world of potential in the medical industry.

Merged ROC and the PR curve

The AUROC value (0.93–1.00) indicated better model performance for the tubband and the considered genes. The AUPRC value (0.99-1.00) for the tubb and, rpoB, InhA, katG and, gyrA/gyrB genes, indicates perfect precision and recall, while the genes rrs, and eis values (0.50), which cannot classify these respective genes, model performance at ICMR-NIRT. Moreover, the AUROC value (0.98-1.00) demonstrated the enhanced model performance of the discussed genes. The AUPRC values were 1.00 gyrA/gyrB and eis genes, which indicated perfect precision and recall, while the tubband and the genes rpoB, katG, InhA, rrs and eis ranged between (0.98-0.50), which indeed indicated better model performance. The value was 1.00 for the tubb and, indicating perfect discrimination, while the rpoB, katG, InhA, gyrA/gyrB, rrs, and eis ranges (0.88–0.96) represented a better performance model. The AUPRC value ranged from 0.83-0.95 for the rpoB, katG, InhA, gyrA/gyrB, rrs, and eis except the tubband (1.00), which exhibited better model performance for phase III.

Unlike conventional AI systems such as CAD4TB for chest radiographs or CNN-based models for smear microscopy, our system addresses a specific void: the automated analysis of molecular tests using LPA strips. While previous investigations have underscored AI's usefulness in radiology and microscopy, scarce attention has been given to strip-format genetic tests. Thus, our research augments current AI diagnostics by targeting a crucial yet insufficiently automated phase in the drug resistance identification process. Human-in-the-loop review was initiated for 37 per cent of FL-LPA and 32 per cent of SL-LPA strips within the validation dataset; among them, only a small fraction (3–9%) of the total strips called for human correction. Faint bands or visual distortions were the main causes of these overrides, which emphasises the system's durability even as essential quality control is upheld through expert evaluation ¹¹.

The model performs reliably with 72 dpi resolution; however, variability in scanners and imaging conditions may affect consistency. Misclassification can occur due to strip misalignment or artifacts. Additionally, training on a specific dataset may limit generalisability, highlighting the need for broader validation. Successful deployment will require compatible scanners (≥72 dpi) and reliable computing infrastructure, standardised staff training for AI use and override procedures, and adherence to national regulatory and data governance frameworks. Addressing these factors is essential for safe, equitable, and sustainable integration into TB program workflows. Moreover, the system will ensure privacy, mitigate bias, and maintain accountability through anonymised data, diverse training, monitoring, and human-in-the-loop oversight.

Despite its limitations in terms of manual interface, the background analysis has proven AI's overall performance to be superior when manual intervention is applied. A key characteristic of this model is its efficiency in reducing interpretation time and delivering reliable results across states. Future research is expected to predict algorithm optimisation without a manual interface.

Using AI solutions reduces interpretation time and enhances the consistency of the interpretation process across states. Direct transmission allows public health to access the patient's information and help end TB. In this study, we demonstrated that AI-based systems improve laboratory workflows for patient-wise LPA testing and directly transmit information to designated information systems for public health action. The outcomes of this research yielded an exceptional F1 score, illustrating high levels of precision and recall when comparing the AI tool to the gold standards, suggesting a better fit when used along with the manual interface.