Evaluation of CAD-based Triage for CXR Interpretation During TB Screening

In 2021, the World Health Organization released guidelines which recommended computer-aided detection (CAD) software as a replacement for radiologists during chest X-ray (CXR) screening for TB.[1] However, clinical workflows which position CAD software as a decision support tool for radiologists, with the aim of improving CXR interpretation accuracy and/or efficiency, may prove to be a more acceptable use case with radiologists. CAD software are now being integrated into breast [2], prostate [3], and lung [4] cancer screening programs in high-income countries in these ways. Yet, there is currently a dearth of literature evaluating CAD software during CXR screening for TB under such use cases.

Freundeskreis Für Internationale Tuberkulosehilfe e.V. (FIT), in collaboration with local public-sector partners, will organize 80 community-based CXR screening events for TB [5,6] across three provinces in Southern Viet Nam (Ba Ria - Vung Tau, Ho Chi Minh City and Long An) as part of a pragmatic clinical trial designed to assess the real-world impact a CAD software deployment. INSIGHT CXR CAD software (Lunit, South Korea) will be used to support CXR interprtation at half of the screening events (randomly selected) by automating the identification of normal CXR images before an on-site radiologist makes a final CXR interpretation (CAD-based triage use case). The other screening events will use only an on-site radiologist for CXR interpretation (usual care).

A retrospective assessment of comparing radiologist only CXR interpretation to CAD-based triage with INSIGHT CXR software showed that CAD-based triage resutled in a -68.9% reduction in human workloads, a -30.1% decrease in CXR abnormality rates and follow-on diagnostic testing, and just a -0.2% reduction in TB detection. This pragmatic clinical trial is neseted within a community-based CXR screening initiative whose scale has been determined by the availability of donor funding. However, a sufficient number of participants will be recruited and screened in each arm to detect a 30% difference (12.4% vs 17.8%) in the proportion of CXR images labelled as abnormal (primary aim).

Study Arms

1. On-site radiologist / usual care (40 screening events; 12,000 participants): All participants in this arm will be screened by CXR. All CXR images will be interpreted by only an on-site radiologist. Participants with an abnormal CXR result from the radiologist will be indicated for diagnostic testing.

2. CAD-based triage / experimental care (40 screening events; 12,000 participants): All participants in this arm will be screened by CXR. All CXR images will first be processed with the INSIGHT CXR CAD software (Lunit, South Korea) to identify the totally normal/clear CXR images; only those with the possibility of containing an abnormality (abnormality score ≥ 20) will be sent to the on-site radiologist. Participants with an abnormal CXR result from the radiologist will be indicated for diagnostic testing.

Aims

1. Compare the difference in the proportion of CXR images which are declared as abnormal by the on-site radiologist between the study arms

2. Compare the difference in the proportion of people diagnosed with TB using the Xpert MTB/RIF Ultra assay among those screened by CXR between the study arms