Artificial intelligence on chest X-ray for tuberculosis screening in Tanzania: a multicentre evaluation.
Authors
Affiliations (6)
Affiliations (6)
- Department of Biomedical Engineering, Muhimbili University of Health and Allied Sciences, Dar es Salaam, Tanzania. [email protected].
- Emerging Technologies for Health Research and Development Laboratory (ETH), Department of Biomedical Engineering, Muhimbili University of Health and Allied Sciences, Dar es Salaam, Tanzania. [email protected].
- Department of Biomedical Engineering, Muhimbili University of Health and Allied Sciences, Dar es Salaam, Tanzania.
- Emerging Technologies for Health Research and Development Laboratory (ETH), Department of Biomedical Engineering, Muhimbili University of Health and Allied Sciences, Dar es Salaam, Tanzania.
- Muhimbili National Hospital Mloganzila, Dar es Salaam, Tanzania.
- School of Computing, University of Derby, Derby, Derbyshire, England.
Abstract
Tanzania has adopted artificial intelligence (AI)-assisted chest X-ray screening for tuberculosis (TB), including the use of CAD4TB version 6, which is registered by the Tanzania Medicines and Medical Devices Authority (TMDA). While GeneXpert, practical reference standard used in routine practice, remains the primary bacteriological confirmatory test in routine practice, there is currently no established national threshold for CAD4TB use in either active case finding (ACF) or passive case finding (PCF) settings. This study evaluates the implementation and operational use of CAD4TB version 6 within mobile TB screening units in Tanzania and highlights challenges affecting its effective use. We conducted a retrospective analysis of screening data from 11,923 individuals collected from mobile clinics equipped with digital X-ray, CAD4TB version 6, and GeneXpert systems. Comparisons were made between manual chest X-ray interpretation, CAD4TB scores, and GeneXpert results within the subset of individuals who underwent confirmatory testing. The findings reveal substantial inconsistencies in screening workflows, including non-uniform use of CAD4TB prior to GeneXpert testing, missing radiological records, and deviations from intended protocols across sites. Descriptive analysis showed that CAD4TB scores generally aligned with GeneXpert-positive cases within the tested subset; however, due to selective application of GeneXpert and incomplete data, these observations cannot be interpreted as measures of diagnostic accuracy. This study should be interpreted as an implementation and operational assessment of AI-assisted TB screening rather than a diagnostic accuracy or threshold-setting study. The findings highlight important gaps in protocol adherence, data completeness, and workflow standardization, underscoring the need for prospective, protocol-driven studies to establish validated national thresholds for CAD4TB use in Tanzania.