The results are in: Artificial intelligence outperforms humans at reading chest x-rays for signs of TB

Computer-aided detection (CAD) products use AI to read X-ray images and predict the likelihood that TB-related signs are present to inform diagnostic decision-making. In theory, and increasingly in practice, AI can be used in synergy with existing human resources to accelerate TB case detection on the ground, modernizing TB programs.

This independent evaluation demonstrated that all evaluated AI products outperformed experienced radiologists. All products were able to halve the number of necessary follow-on diagnostic tests while retaining high sensitivity of above 90%. Even when reducing the number of follow-on tests by two-thirds, all AI products were more than 80% sensitive. Two products also met the aspirational 90% sensitivity and 70% specificity target product profile set by the World Health Organization (WHO) for a TB triage test. AI could therefore enable TB programs to reduce costs without compromising dramatically on the number of cases detected.

The research, led by Stop TB Partnership experts, rigorously evaluated the most recent versions of five on-the-market CAD products. The evaluation tested AI on an external dataset of 23,954 chest X-rays collected from TB screening clinics run by collaborator icddr,b in Dhaka, Bangladesh. The reading result from AI was compared against Xpert, the bacteriological reference standard.

The Bangladeshi screening program was funded by the Stop TB Partnership’s TB REACH initiative. The evaluation tested CAD4TB (Delft Imaging Systems, the Netherlands), InferRead DR Chest (Infervision, China), JF CXR-1 (JF Healthcare, China), Lunit INSIGHT CXR (Lunit, South Korea), and qXR (Qure.ai, India). This is the first-ever independent evaluation of Infervision and JF Healthcare’s products in peer-reviewed literature.

In March 2021, evidence produced by TB REACH contributed to the update of WHO policy which now, for the first time, recommends the use of CAD for TB triage and screening. Products in the Bangladeshi study, not previously evaluated, were found to perform to the high standard required by the updated WHO policy. This demonstrates the potential of CAD products to revolutionize case detection efforts by TB programs, especially where radiologists are scarce.

Not stopping there, as new products and software emerge, TB REACH continues to lead their independent evaluation, utilizing its large online archive of chest X-rays to promote evidence-based implementation. Meanwhile, Stop TB Partnership’s Digital Health Technology Hub leverages cross-partnership expertise to support implementers who wish to use the newest digital tools, including AI.

“It is very good,” said Dr. Lucica Ditiu, Executive Director of the Stop TB Partnership. “The Stop TB Partnership continues to pave the way for ground-breaking technologies and innovations in the fight against TB. In our collective mission to end TB, we need to change our mindsets and use all new tools and guidance for those that need them most.”

Read the full study here.

Source: Stop TB Partnership