AN AI-powered lung ultrasound tool has outperformed human experts in diagnosing pulmonary tuberculosis (TB), offering a fast, accessible, and sputum-free solution that could transform TB triage in high-burden, resource-limited settings.
Presented at ESCMID Global 2025, the study evaluated the ULTR-AI suite, an integrated set of deep learning models developed to interpret lung ultrasound scans in real time using portable, smartphone-connected devices. Conducted in a tertiary centre in Benin, West Africa, the study included 504 patients after exclusions, 38% of whom had microbiologically confirmed pulmonary TB. Researchers applied a 14-point standardised lung ultrasound protocol and compared AI model performance against human expert readings. A single MTB Xpert Ultra test was used as the reference standard for diagnosis.
The AI model ULTR-AI (max) achieved 93% sensitivity and 81% specificity, with an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.93 (95% CI: 0.92–0.95), surpassing the World Health Organization (WHO)’s minimum target for non-sputum-based TB triage tools (90% sensitivity and 70% specificity).
Compared to human experts, the AI system demonstrated a 9% improvement in diagnostic accuracy. The AI detected both human-recognisable features, such as consolidations and interstitial changes, and more subtle patterns that may be missed by trained clinicians. Importantly, the tool showed promise in identifying early, sub-centimetre pleural lesions associated with TB. The tool’s performance was consistent even in patients with HIV co-infection and those with a prior history of TB. Once integrated into a mobile app, the AI model provided instant diagnostic results at the point of care.
This study highlights the potential of AI-enhanced ultrasound as a frontline diagnostic tool for TB, particularly in regions with limited access to radiological services. With real-time interpretation and minimal operator dependency, ULTR-AI could significantly improve early detection and reduce patient drop-out rates.
Reference
Suttels V et al. Lung ultrasound for the detection of pulmonary tuberculosis using expert and AI-guided interpretation. ESCMID Global 2025, 11-15 April, 2025.
