A study in Dhaka, Bangladesh, has found that artificial intelligence was able to detect tuberculosis better than human radiologists, but the results shouldn’t sow fear that the health professionals may be replaced by the technology in either the near term or the future, according to the head of the Stop TB Partnership.
“There should not be such fear as radiologists will always be needed. Currently, there is such a shortage of radiologists everywhere that, for the time being, there is no concern. In the long run, when AI is really scaled up, there will be a need for proper planning on x-ray operations and reading needs,” Stop TB Partnership Executive Director Lucica Ditiu told Devex in an email.
In the future, this shift could free up radiologists’ time and allow them to focus on different imaging technologies for TB and other diseases, she added.
“These critical findings can provide support for the uptake of the technology to more easily detect people [with TB] who would otherwise have been missed.”
— Lucica Ditiu, executive director, Stop TB PartnershipThe study, conducted by the Stop TB Partnership and health research institute icddr,b, compared chest X-ray readings made by three certified radiologists in Dhaka with those done by five AI algorithms. The readings were then measured against the results of patients’ nucleic acid amplification tests, which were employed as the reference standard for TB diagnosis.
The study showed that all five AI products outperformed the experienced radiologists in detecting TB. However, the findings are limited by the absence of children and HIV testing from the study. As such, the performance of the AI algorithms for young people and those living with HIV requires further research.
Ditiu said the results show that AI could reduce the number of patients requiring expensive diagnostic testing to confirm TB by more than half “while ensuring we don’t miss many people.” All five AI algorithms demonstrated sensitivity above 90%, and even when follow-up diagnostic testing was reduced by two-thirds, the sensitivity of the AI algorithms was still 80% and above.
The results also point to how AI can help screen for TB in places with few trained radiologists, and implementers may use the findings when deciding how to use AI in certain settings, Ditiu said.
“These critical findings can provide support for the uptake of the technology to more easily detect people who would otherwise have been missed and save thousands of dollars in expensive molecular tests on people who do not need them,” she added.
Sign up for Devex CheckUp
The must-read weekly newsletter for exclusive global health news and insider insights.
A number of studies have evaluated the success of AI algorithms in detecting TB via chest X-rays, but there have been few independent evaluations of AI for reading chest X-rays to detect TB. The Dhaka study also adds to the body of evidence comparing the performance of AI against radiologists.
In addition, the findings provide evidence to support the World Health Organization’s recommendation in March for using computer-aided detection — or CAD — software to analyze digital chest X-rays in TB screening and triage among patients ages 15 and above, Ditiu said.
“Many new CAD products are in the pipeline, and new versions of existing products emerge almost annually, so studies like these will continue to be essential to keep a handle on how they perform in real-world datasets and ensure implementers can have confidence when using the tools,” she added.
About the author
Jenny Lei Ravelo@JennyLeiRavelo
Jenny Lei Ravelo is a Devex Senior Reporter based in Manila. She covers global health, with a particular focus on the World Health Organization, and other development and humanitarian aid trends in Asia Pacific. Prior to Devex, she wrote for ABS-CBN, one of the largest broadcasting networks in the Philippines, and was a copy editor for various international scientific journals. She received her journalism degree from the University of Santo Tomas.
