TY - GEN
T1 - TB Screening App
T2 - 1st International Medical Image Computing in Resource Constrained Settings Workshop and Knowledge Interchange, MIRASOL 2025, held in Conjunction with MICCAI 2025
AU - Oladokun, Ajibola S.
AU - Malila, Bessie
AU - Hamada, Yohhei
AU - Lilaonitkul, Watjana
AU - Rangaka, Molebogeng X.
AU - Mutsvangwa, Tinashe E.M.
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026
Y1 - 2026
N2 - Latent tuberculosis infection (LTBI) is the precursor to active TB. Africa, South-east Asia, and Western-Pacific account for about 80% of global TB infections. The tuberculin skin test (TST) is the most common test for LTBI in these regions due to its low cost and simplicity. It involves the subdermal injection of tuberculin which may trigger the formation of an induration 48 to 72 h later. Patients often don’t return to the clinic to read the induration diameter to determine LTBI diagnosis. Researchers in Cape Town, South Africa, developed the TB Screening App to enable automated LTBI diagnosis from home to prevent the need for secondary clinical visits. The app utilized a photogrammetry-based backend which did not yield reliable results. In this paper, we propose a vision transformer-based pipeline to replace the backend of the app and enable accurate LTBI diagnosis. We developed and tested the pipeline to classify 1,026 images collected in previous studies using the app. The pipeline enabled a mean balanced accuracy of 82.3% ± 2.9 for LTBI classification, and produced attention maps that visually correlate with a clinician’s readings. This shows the potential of the pipeline to facilitate automated LTBI screening in low resource settings and eliminate the need for secondary clinical visits by patients.
AB - Latent tuberculosis infection (LTBI) is the precursor to active TB. Africa, South-east Asia, and Western-Pacific account for about 80% of global TB infections. The tuberculin skin test (TST) is the most common test for LTBI in these regions due to its low cost and simplicity. It involves the subdermal injection of tuberculin which may trigger the formation of an induration 48 to 72 h later. Patients often don’t return to the clinic to read the induration diameter to determine LTBI diagnosis. Researchers in Cape Town, South Africa, developed the TB Screening App to enable automated LTBI diagnosis from home to prevent the need for secondary clinical visits. The app utilized a photogrammetry-based backend which did not yield reliable results. In this paper, we propose a vision transformer-based pipeline to replace the backend of the app and enable accurate LTBI diagnosis. We developed and tested the pipeline to classify 1,026 images collected in previous studies using the app. The pipeline enabled a mean balanced accuracy of 82.3% ± 2.9 for LTBI classification, and produced attention maps that visually correlate with a clinician’s readings. This shows the potential of the pipeline to facilitate automated LTBI screening in low resource settings and eliminate the need for secondary clinical visits by patients.
KW - AI
KW - Induration
KW - LTBI
KW - Latent tuberculosis infection
KW - Mobile health
KW - TB Screening App
KW - TST
KW - Tuberculin skin test
KW - Vision transformer
UR - https://www.scopus.com/pages/publications/105042918427
U2 - 10.1007/978-3-032-13654-1_21
DO - 10.1007/978-3-032-13654-1_21
M3 - Conference contribution
AN - SCOPUS:105042918427
SN - 9783032136534
T3 - Lecture Notes in Computer Science
SP - 206
EP - 216
BT - Medical Image Computing in Resource Constrained Settings - 1st International Workshop, MIRASOL 2025, Held in Conjunction with MICCAI 2025, Proceedings
A2 - Anazodo, Udunna
A2 - Raymond, Confidence
A2 - Zhang, Dong
A2 - Kurt, Mehmet
A2 - Lekadir, Karim
A2 - Crimi, Alessandro
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 27 September 2025 through 27 September 2025
ER -