AI helps speed chest X-ray reads to fight tuberculosis in India

India faces a critical shortage of radiologists, which delays interpretation of chest radiographs (CXRs) and can allow lung and heart diseases to progress before treatment begins. Those delays contribute to higher morbidity and mortality, particularly for high-burden conditions such as tuberculosis. To address this gap, researchers led by Vinit Singh tested an artificial intelligence tool called Lenek Intelligent Radiology Assistant (LIRA) as a possible screening and triage solution. Rather than replacing clinicians, the study evaluated whether LIRA could reliably flag abnormalities and priority cases from routine CXRs so patients could be seen faster. The work was retrospective and multi-source: de-identified chest radiographs from geographically diverse institutions were fed into LIRA, and the tool’s outputs were compared to established ground truth reports. The goal was to determine how well LIRA detects general abnormalities and specific pathologies—most notably tuberculosis—but also consolidation, pleural effusion, pneumothorax, and cardiomegaly. By establishing accuracy across different datasets, the researchers aimed to understand whether LIRA could be a consistent adjunct to human readers in settings with limited radiology capacity.

The study carried out a retrospective multi-source validation of the diagnostic accuracy of Lenek Intelligent Radiology Assistant (LIRA). De-identified chest radiographs were processed by LIRA models and the resulting interpretations were compared against established ground truth reporting. The team calculated sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals across varying probability thresholds for each pathology. Results showed strong performance for general abnormality detection with AUROC between 0.93 and 0.986, sensitivity ranging 84.4–97.1%, and specificity 88.9–92.4%. For tuberculosis triaging, two datasets labeled Shenzhen & Montgomery produced sensitivities of 88.5–89.7% and specificities of 89.9–90.5%, while the Jaypee dataset showed 98.7% sensitivity and 63.6% specificity. Consolidation detection had AUROC 0.884–0.895 with 96.4–96.9% sensitivity and 70.8–77.1% specificity. Pleural effusion detection showed AUROC 0.942–0.967 with sensitivity 79.7–99.1% and specificity 81.2–87.7%. Pneumothorax had AUROC 0.87 with 90.6–94.8% sensitivity and 79.5–82.7% specificity, and cardiomegaly detection had AUROC 0.883 with 95.1% sensitivity and 81.6% specificity.

The study’s findings indicate that LIRA delivers consistent diagnostic performance across a range of chest pathologies and across radiographs from diverse geographic sources. Its particular strengths were in detecting general abnormalities and in tuberculosis screening, where high sensitivity helps ensure fewer cases are missed during initial triage. The model’s risk-stratified outputs mean it can prioritize likely positive cases for faster human review, which is especially valuable where radiologist availability is limited. By reducing turnaround times for CXR interpretation, LIRA could help clinicians begin further tests or treatment sooner, potentially limiting disease progression and lowering morbidity. The authors conclude that LIRA is a reliable adjunct solution—intended to support, not replace, radiologists—and that its consistent accuracy supports use as part of efforts to reduce diagnostic delays and to back India’s tuberculosis elimination goals. Further real-world deployment and monitoring would be needed to measure impact on care pathways and patient outcomes.