Deep learning of point-of-care lung ultrasound and clinical data for pulmonary tuberculosis triage in Uganda

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Rationale

Point-of-care lung ultrasound is portable and radiation-free and could extend tuberculosis screening to settings without chest radiography, but no standardized criteria exist for interpreting it for tuberculosis.

Objectives

To develop and internally evaluate deep learning models that classify pulmonary tuberculosis from lung ultrasound images, alone and combined with routine clinical variables, and to derive an abbreviated scanning protocol.

Methods

In a prospective study in Kampala, Uganda (2023–2025), we enrolled people aged 13 years or older with microbiologically confirmed pulmonary tuberculosis, their household contacts, and symptomatic people with other respiratory conditions. One still frame from each of 12 lung ultrasound zones was analyzed with an attention-based multiple-instance learning network. We compared clinical-only (random forest), ultrasound-only, and combined models and derived a 6-zone protocol from training-set attention weights. Participants were allocated to training (n = 223), validation (n = 20), and held-out test (n = 73) sets.

Results

Of 318 enrolled participants, 316 were analyzed (118 with and 198 without tuberculosis). In the test set, areas under the receiver operating characteristic curve were 0.88 (95% confidence interval, 0.80–0.95) for the clinical-only model; 0.93 (0.86–0.97) and 0.95 (0.89–0.99) for the 12-zone ultrasound-only and combined models; and 0.91 (0.83–0.97) and 0.93 (0.86–0.99) for the corresponding 6-zone models. At 90% sensitivity, specificity was 82.2% (34.9–97.8) for the 12-zone combined model and 91.1% (25.0–98.0) for the 6-zone combined model. Point estimates for all 12-zone models and the 6-zone combined model met the 2025 World Health Organization minimum accuracy targets for a high-sensitivity screening test (≥90% sensitivity, ≥60% specificity).

Conclusion

Lung ultrasound analyzed with attention-based deep learning and combined with clinical variables showed high internal discrimination for pulmonary tuberculosis, and a simplified 6-zone protocol retained most of this performance. External validation in intended-use screening populations is needed.

Article activity feed