Research map / Architectures and efficiency / RNNs and normalization
Automatic Interpretation of Line Probe Assay Test for Tuberculosis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an automated computer-vision and ML pipeline (DETR band detection, BiLSTM band classification, hierarchical neural network drug interpretation, TrOCR lab-ID recognition, and a human-in-the-loop rejection module) to interpret Line Probe Assay strips for drug-resistant TB.
Problem
Manual interpretation of Line Probe Assay (LPA) strips for drug-resistant tuberculosis is slow, labor-intensive, and requires expert microbiologists, delaying treatment decisions.
Approach
Decomposes LPA interpretation into strip extraction, band detection with DETR, sequence-based band classification with a BiLSTM over centroid distances to reference bands, and a hierarchical network mirroring the deterministic clinical rules, plus a rejection/discrepancy/fidelity module that routes ambiguous strips to lab technicians; simulated sheets are used for pretraining.
Key results
- HNN drug interpretation reaches 94% (RMP, FL) and 88-98% accuracy across SL drugs, versus 66-81% for deterministic rules (DET).
- DETR band detection achieves 89.27 FL mAP and 78.2 SL mAP, versus 57.5 and 49.7 for Faster R-CNN.
- With human-in-the-loop review, accuracy exceeds 98% for all drugs in FL and SL, with only about 4% of strips requiring intervention for RMP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FRCNN with DETR for band detection | LPA band detection · FL and SL LPA strips | mAP: 57.5 to 89.27 (FL); 49.7 to 78.2 (SL) | Helped |
| Use DETR+BiLSTM instead of FRCNN+BiLSTM for band classification | band classification · FL and SL LPA strips | sensitivity and specificity: SN 0.87 to 0.97 (FL); SN 0.83 to 0.96 (SL) | Helped |
| Replace deterministic rules (DET) with hierarchical neural network (HNN) for drug interpretation (ablation) | clinical interpretation · FL and SL drugs (Table 7) | drug interpretation accuracy (%): RMP FL 76 to 94; INH FL 81 to 94; FLQ SL 66 to 88; KAC SL 62 to 95; LLK SL 76 to 98 | Helped |
| Pretrain TrOCR on IAM handwriting dataset before fine-tuning for lab ID recognition (ablation) | TrOCR lab ID recognition · Lab ID dataset | accuracy: 0.92 to 0.943 | Helped |
| Use TrOCR instead of custom CNN-LSTM for lab ID recognition (ablation) | lab ID recognition · Lab ID dataset | accuracy: 0.84 to 0.92 | Helped |
| Add human-in-the-loop review of flagged ambiguous strips (ablation) | HNN drug interpretation for RMP · FL gold standard dataset | accuracy (%): 94 to 98 with 37% of strips sent for review | Helped |
| Human-in-the-loop review | all drugs · FL and SL gold standard dataset | accuracy (%): Accuracy above 98% on all drugs with a small fraction of strips reviewed | Helped |
Surprising findings
- Improvements from human-in-the-loop review are attributed to incompressible label noise rather than more data or better models.
Limitations
- Rare band combinations and resistive SL drugs (KAC, LLK) have very low support, so estimates are unreliable.
- Light or faint MUT bands are a major source of error, and a single band error can change drug outcomes.
- Generalization to rare unseen band configurations is not guaranteed given limited dataset diversity.
Experimental setup
Datasets: Real LPA sheet dataset (1200 sheets, 13482 FL and 1746 SL strips) Simulated LPA sheets (~10K) IAM handwriting dataset
Benchmarks: Real LPA strip test split (FL and SL) Gold standard LPA evaluation dataset
Metrics: accuracy precision recall mAP@0.5 sensitivity specificity
Baselines: Faster R-CNN CNN-LSTM DET (deterministic rules) TrOCR pre-trained on IAM
Builds on: DETR TrOCR BiLSTM focal loss
Closest papers by meaning
- Deep Reinforcement Learning for Early Diagnosis of Lung Cancer (AAAI 2024 · similarity 0.75)
- Optimizing Gene-Based Testing for Antibiotic Resistance Prediction (AAAI 2025 · similarity 0.74)
- ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening (NeurIPS 2026 · similarity 0.74)
- A Trusted Lesion-assessment Network for Interpretable Diagnosis of Coronary Artery Disease in Coronary CT Angiography (AAAI 2025 · similarity 0.74)
- Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning (ICML 2025 · similarity 0.74)
- Toxicity Detection for Free (NeurIPS 2024 · similarity 0.74)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.