What if nurses could screen everyone for TB, not just people with symptoms?
A new paper by Christiaan Guldenhuys and co-authors from the CAD LUS4TB project shows how this could work. Transferring echocardiography pretraining to lung ultrasound for tuberculosis screening was presented at the Southern Africa Telecommunication Networks and Applications Conference (SATNAC), held 11–15 October 2026. The two-minute video explains how it works.
In many primary-care clinics, TB screening still relies on a symptom questionnaire, and people without obvious symptoms are sent home. CAD LUS4TB is developing machine learning techniques for handheld lung ultrasound so that nurses can screen everyone and refer only those who need a confirmatory test.
Labelled lung ultrasound data is scarce, so we tested whether a model pretrained on heart ultrasound helps. On held-out patients, at the WHO target of 90% sensitivity, specificity rose from 60% to 79%. That means potentially ~170 fewer false positive referrals for every 1,000 people screened. Unfortunely heart pretraining did no better than pretraining on everyday video. But due to the limited labelled data it can't be said for certain.