Chest X-ray interpretation using AI can detect more TB cases

India has a powerful technology — AI-assisted chest X-ray — to screen presumptive TB cases. The AI algorithm (qXR), which was developed by the Mumbai-based Qure.ai, can help in early detection of people with presumptive TB disease in less than a minute, including people with subclinical TB. As per the 2019-2021 National TB prevalence survey in India report, nearly 43% of TB cases would have been missed if a chest X-ray was not included.

When used at scale for population-based screening or at least for targeted screening, the AI software combined with molecular tests for TB disease confirmation can vastly increase detection rates. Systematic screening for TB disease for early diagnosis is one of the main End TB strategies.

The qXR software got the European CE certification early this year. The Indian drug regulator too has cleared it a few months ago. The qXR was one of the three AI algorithms that the WHO had referenced to when updating the TB screening guidelines in March 2021. In 2021, WHO recommended the use of CAD products which use the AI to automate interpretation of digital chest X-rays for TB screening and triage in people older than 15 years.

Awaiting a nod

Though there are no restrictions on State governments to use the software for TB detection, the Central TB Division’s directive to States to procure and deploy the tool for early TB disease diagnosis will go a long way in widespread adoption. Since States look up to the Central TB Division, the absence of a policy guidance has led to low adoption of the powerful tool. The Central TB Division is yet to recommend the use of qXR as it is waiting for an approval from the Health Technology Assessment, which has evaluated three AI software, including qXR.

The qXR software meets the WHO requirement with over 90% sensitivity and more than 70% specificity in people older than 15 years. In a large study (The Lancet Digital Health) involving nearly 24,000 people aged over 15 years done in Bangladesh to evaluate the use of AI in detecting TB from chest X-rays, qXR had 90.2% sensitivity and 74.3% specificity. A 2019 study (Scientific Reports) found qXR to have over 95% sensitivity and over 80% specificity.

The 2019 study demonstrated that qXR has the “potential to increase capacity and aid TB diagnosis, especially in settings with a shortage of trained human readers” which is a huge shortcoming in using chest X-rays for TB screening.  

India’s ambitious goal of “eliminating” TB by 2025 can become remotely possible only when early diagnosis and initiation of care for millions of people with TB becomes a reality. Large-scale use of AI-assisted chest X-rays for screening is the first crucial step in the TB care cascade to achieve this goal.

In the Ca Mau Province of Vietnam, a community-wide screening of people older than 15 years using a molecular test between 2014 and 2017 resulted in a lower prevalence of pulmonary TB in 2018 than standard passive case detection alone. Unlike in Vietnam, the use of the AI algorithm to read digital X-rays prior to molecular testing as part of community screening will not only reduce TB prevalence but also greatly minimise the number of molecular tests needed to detect TB cases.

The qXR algorithm is already being used in over 50 countries. In India, 24 States are using it in about 150 sites. “But nowhere in India is the software used at scale. And in some States, the qXR software is used in just one site, like in Kerala,” says Dr. Shibu Vijayan, Medical Director-Global Health at Qure.ai. “A digital solution that is not at scale will not make any impact.” While around six outreach mobile vans in Chennai use qXR for TB case detection, nine government hospitals and one mobile van in Mumbai are equipped with the AI software for TB surveillance. The AI algorithm was first installed in January 2021 at the S.K. Patil Hospital in Mumbai, and a month earlier (December 2020) in a mobile van.

Encouraging results

As of December 2022, over 1,00,000 X-rays from the nine Mumbai hospitals were screened for TB. “This is the count of all X-rays taken in Mumbai public hospitals, and we were asked to scan them for TB,” says Dr. Vijayan. “It is difficult to pinpoint who is refereed for what. However, we did an estimation assessment in a few select facilities where we found 30% of people diagnosed with TB were not from TB referral X-rays. These were routine X-rays taken for purposes other than TB diagnosis thus making it possible to detect TB in non-presumptive TB cases.” The Mumbai exercise helped in detecting 13% additional TB cases. And the use of the algorithm to screen the X-rays increased the yield (positivity) of GeneXpert molecular tests by 18-27%.

In the case of the mobile van using the qXR algorithm, of the 10,000 people screened for TB, one in four presumptive TB cases identified by the algorithm turned out to be a confirmed TB case.

“The AI algorithm for screening presumptive TB cases has been quite useful. It made incidental diagnosis of TB disease possible in people who had come for completely different purposes,” says Dr. Mangala Gomare, former Executive Health Officer at Brihanmumbai Municipal Corporation (BMC), Mumbai

Active case finding got a boost early this year when Qure.ai partnered with Mylab Discovery Solutions to use the qXR software in Mylab’s portable chest X-ray device (MyBeam). This will enable screening of presumptive TB cases with the AI algorithm even in rural areas. “Our portable device cuts the amount of X-ray exposure to 1/20th to 1/30th of a normal X-ray even while capturing all the details. This is because our device is tailormade to image only the chest,” says Saurabh Gupta, Head of Digital and Public Health at Mylab, Pune.

Besides digital X-rays, the qXR algorithm can also be used for detecting TB disease from film-based X-rays. To do this, a photograph of the X-ray film is taken and the digital image is processed by the algorithm just like it does with a digital X-ray. “Our internal validation confirmed that for AI function, there is no significant difference in impact between digital and analogue [film-based] X-rays for TB detection,” Dr. Vijayan says.

The biggest advantage of using the AI algorithm even on film-based X-rays that have been turned into digital data and if adopted in all the 730 district hospitals in India is the probability of detecting around 80,000 incidental TB cases from around 2,80,000 chest X-rays taken each year for reasons other than TB diagnosis.

While WHO’s guidelines for AI-assisted detection of TB using digital X-rays is restricted to people above 15 years, Qure.ai’s AI algorithm has a CE certification for use in the paediatric population. The qXR software was successfully used in Bangladesh in children above three years and in children above four years in Myanmar. “Our algorithm for paediatric population has been independently validated by the STOP TB Partnership; the study is expected to be published soon,” Dr. Vijayan says.

Currently paediatric X-rays are being cropped into adult size before using the AI algorithm to flag presumptive TB cases. Unlike in adults, paediatric pulmonary TB detection using X-rays is a challenge as radiological evidence of pulmonary TB in children is less specific. “We are trying to come up with more paediatric-specific algorithms with disease severity probability,” he says.