Speaker's Highlight

• Don Sin University of British Columbia, St. Paul Hospital (Canada)
  Kenneth R. Chapman Toronto General Hospital Research Institute (Canada)
• Parameswaran Nair McMaster University (Canada)
  Carolyn Calfee UCSF (U.S.A.)
• Gregory P. Downey University of Colorado School of Medicine (U.S.A.)
  David A. Schwartz University of Colorado School of Medicine (U.S.A.)
• Neil Schluger Tuberculosis Control Branch, California Department of Public Health (U.S.A.)
  Nick Kim Critical Care & Sleep Medicine, University of California San Diego (U.S.A.)
• Nicola Hananiah Baylor College of Medicine (U.S.A.)
  Jae-Joon Yim Seoul National University College of Medicine (Republic of Korea)
• Koichiro Asano Tokai University School of Medicine (Japan)
  Diahn-Warng Perng Taipei Veterans General Hospital (Taiwan)
• Konstantinos Kostikas University of Ioannina (Greece)
  Karin Klooster University Medical Center Groningen (Kingdom of the Netherlands)

• AI-assisted chest radiography for tuberculosis programs

  Artificial intelligence is changing how clinicians read medical images, and few areas feel that shift more directly than tuberculosis screening. A plain chest radiograph remains one of the cheapest, fastest, and most widely available tools for spotting pulmonary disease, yet human interpretation is variable and access to trained radiologists is uneven.

  In countries with a low background incidence of TB, the case for AI-supported reading is shifting from case-finding toward targeted screening of high-risk groups. Australia sits firmly in that category. National notifications hover around 1,400 to 1,600 cases per year, with disproportionately higher rates among Aboriginal and Torres Strait Islander peoples and among people born overseas in high-burden regions.

  Digital chest X-ray systems, paired with cloud-based inference, are now portable enough to reach clinics in remote Queensland, the Torres Strait, and the Kimberley region of Western Australia. The technology is moving from research curiosity to operational tool, reshaping how TB programs plan their diagnostic pathways.

  The global burden of TB and why imaging matters

  Despite decades of progress, tuberculosis still kills more than a million people worldwide each year. The World Health Organization has endorsed computer-aided detection of TB on chest radiographs as a triage or screening aid, particularly where bacteriological confirmation is hard to obtain or where laboratory capacity is limited.

  Cough, fever, and weight loss are the textbook symptoms, but in early or extra-pulmonary disease the chest film is often the first objective evidence a clinician sees. In the Western Pacific Region, which includes Australia and several neighbouring high-burden countries, the WHO has emphasised active case-finding as a pillar of elimination strategy.

  For Australian clinicians the practical challenge is not finding thousands of cases. It is finding the few that exist, often in communities that already face barriers to specialist care. An algorithm that flags suspicious films for priority review can stretch a limited radiology workforce and shorten time to treatment.

  How deep learning reads a chest X-ray

  Modern systems rely on convolutional neural networks trained on tens or hundreds of thousands of labelled radiographs. The model learns to associate patterns of consolidation, cavitation, hilar enlargement, and upper-lobe infiltrates with a TB label, then outputs a probability score or a heatmap overlay.

  Most commercially available tools were developed using datasets from South Africa, India, and parts of Eastern Europe. External validation matters: performance drops when the training population differs substantially from the target population in terms of disease prevalence, co-morbidities such as HIV or diabetes, or even X-ray equipment manufacturer.

  Researchers at the University of Sydney and the CSIRO's Australian e-Health Research Centre in Brisbane have begun testing these models on local archives, partly to benchmark performance and partly to explore whether re-training with local data closes the gap. Early findings suggest calibration, not just discrimination, is the harder problem to solve.

  Validation studies and real-world performance

  Published evidence has grown quickly. A systematic review pooled data from more than thirty evaluations and found pooled sensitivity in the 90s percent range and specificity in the 80s, though confidence intervals were wide and study quality variable. Headline numbers can mislead when prevalence is low.

  In practice, an AI threshold can be tuned to favour sensitivity or specificity depending on context. For mass screening in a high-burden setting, a sensitive threshold catches more cases at the cost of more follow-up visits. For a low-incidence country like Australia, a more specific threshold reduces unnecessary sputum collection and clinic visits.

  A small pilot at a refugee health clinic in Melbourne reported that an AI triage system reduced radiologist reporting turnaround from several days to under twenty-four hours for flagged films. While the sample was modest, the workflow benefit was tangible and reproducible across different radiologists.

  Implementation challenges in diverse healthcare settings

  Technology is rarely the bottleneck. Licensing costs, integration with picture archiving systems, data governance, and clinician trust all shape whether an AI tool actually reaches a patient. In rural New South Wales and across the vast catchments covered by services such as the Royal Flying Doctor Service, bandwidth and offline capability can be just as important as raw accuracy.

  Workforce culture matters too. Senior clinicians in tertiary centres may be sceptical of an algorithm that disagrees with their read, while junior doctors may over-rely on a score without reviewing the film themselves. Clear positioning of the tool as a second reader, not a replacement, has helped adoption in early Australian deployments.

  Related respiratory work, such as exhaled nitric oxide monitoring, shows how biomarker-based decision support is being integrated into clinical workflows more broadly. The pattern is similar: an objective signal that complements clinical judgement rather than replaces it.

  Practical considerations for sites considering adoption include:

  • Mapping the clinical pathway before selecting a tool, so the AI fits an actual workflow gap
  • Confirming the vendor's training data includes demographics similar to the local population
  • Building a local audit loop that flags drift in performance over months, not just at go-live
  • Budgeting for human review of every flagged film, since no current system is autonomous

  Looking ahead: integration, regulation, and equity

  The next wave of work is less about the algorithm itself and more about how it lives inside a health system. Software-as-a-medical-device regulation in Australia is overseen by the Therapeutic Goods Administration, and updated guidance now explicitly addresses adaptive AI and continuous-learning systems. Vendors will need to demonstrate ongoing post-market surveillance, not just initial clearance.

  Equity is the harder question. If an algorithm performs best on the populations represented in its training set, then the people most likely to be missed are those already underserved. Aboriginal Medical Services in the Northern Territory and primary care clinics serving recently arrived migrants are obvious partners for prospective evaluation, not afterthoughts.

  Key questions for any procurement decision include:

  • Who is accountable when the algorithm and the radiologist disagree
  • How the tool will perform on paediatric and pregnant patients, who are often under-represented in training sets
  • Whether the vendor will share model cards and failure modes openly
  • How local data will be handled, stored, and protected under Australian privacy law

  The clearest lesson from early adopters is that AI for chest radiograph interpretation is most valuable when it gives time back to clinicians and earlier answers to patients. For low-incidence settings like Australia, that means smarter screening of high-risk groups, faster triage in busy emergency departments, and stronger support for remote and Indigenous health services that already do extraordinary work with limited resources.

  Richard Russell Nuffield Department of Clinical Medicine, University of Oxford (United Kingdom)
• Mona Bafadhel King’s College London (United Kingdom)
  David Jackson Guy’s and St Thomas’ Hospital, King’s College London (United Kingdom)
• James Chalmers University of Dundee (United Kingdom)
  David Price University of Aberdeen (United Kingdom)