AI-powered lung ultrasound outperforms human experts in diagnosing tuberculosis

A pioneering study presented today at ESCMID Global 2025 has demonstrated that an AI-powered lung ultrasound outperforms human experts by 9% in diagnosing pulmonary tuberculosis (TB).

The ULTR-AI suite analyses images from portable, smartphone-connected ultrasound devices, offering a sputum-free, rapid, and scalable alternative for TB detection. The results exceed the World Health Organization (WHO) benchmarks for pulmonary tuberculosis diagnosis, marking a major opportunity for accessible and efficient TB triage.

Despite previous global declines, TB rates rose by 4.6% from 2020 to 2023.² Early screening and rapid diagnosis are critical components of the WHO's 'End TB Strategy,' yet many high-burden countries experience substantial patient dropouts at the diagnostic stage due to the high cost of chest x-ray equipment and a shortage of trained radiologists.³  

These challenges underscore the urgent need for more accessible diagnostic tools. The ULTR-AI suite leverages deep learning algorithms to interpret lung ultrasound in real time, making the tool more accessible for TB triage, especially for minimally trained healthcare workers in rural areas. By reducing operator dependency and standardising the test, this technology can help diagnose patients faster and more efficiently."

Dr. Véronique Suttels, lead study author

The ULTR-AI suite comprises three deep-learning models: ULTR-AI predicts TB directly from lung ultrasound images; ULTR-AI (signs) detects ultrasound patterns as interpreted by human experts; and ULTR-AI (max) uses the highest risk score from both models to optimise accuracy.

The study was conducted at a tertiary urban centre in Benin, West Africa. After exclusions, 504 patients were included, with 192 (38%) confirmed to have pulmonary TB. Among the study population, 15% were HIV-positive and 13% had a history of TB. A standardised 14-point lung ultrasound sliding scan protocol was performed, with human experts interpreting images based on typical lung ultrasound findings. A single sputum molecular test (MTB Xpert Ultra) served as the reference standard.

ULTR-AI (max) demonstrated 93% sensitivity and 81% specificity (AUROC 0.93, 95% CI 0.92-0.95), exceeding WHO's target thresholds of 90% sensitivity and 70% specificity for non-sputum-based TB triage tests.

"Our model clearly detects human-recognisable lung ultrasound findings-like large consolidations and interstitial changes-but an end-to-end deep learning approach captures even subtler features beyond the human eye," said Dr. Suttels. "Our hope is that this will help identify early pathological signs such as small sub-centimetre pleural lesions common in TB."

"A key advantage of our AI models is the immediate turnaround time once they are integrated into an app," added Dr. Suttels. "This allows lung ultrasound to function as a true point-of-care test with good diagnostic performance at triage, providing instant results while the patient is still with the healthcare worker. Faster diagnosis could also improve linkage to care, reducing the risk of patients being lost to follow-up."