Skip to main content
Dryad

Comparison of different Lunit INSIGHT CXR software versions when reading chest radiographs for tuberculosis

Codlin, Andrew1; Vo, Luan1; Dao, Thang1; Forse, Rachel1; Dang, Ha T. M.2; Nguyen, Lan H.2; Nguyen, Hoa B.3; Dinh, Luong V.3; Sidney Annerstedt, Kristi4; Lundin, Johan4; Lonnroth, Knut4

Published Mar 14, 2025 on Dryad. https://doi.org/10.5061/dryad.5tb2rbpcv

Data files

Mar 14, 2025 version files 20.85 GB

Click names to download individual files Select up to 11 GB of files for zip download

Abstract

New versions of computer-aided detection (CAD) software for chest X-ray (CXR) interpretation during tuberculosis (TB) screening are regularly released, which purport to have incremental performance gains. We independently measured the differences in CAD software performance between INSIGHT CXR (Lunit, South Korea) versions 3.1.0.0 and 3.9.0.1. A well-characterized Digital Imaging and Communications in Medicine (DICOM) test library was compiled using data from a community-based TB screening initiative in Ho Chi Minh City, Viet Nam. The performance of Lunit CAD software versions was compared by measuring the area under the receiver operating characteristic curve (AUC), stratified by key clinical and demographic variables and using Xpert MTB/RIF Ultra (Ultra) test results as the reference standard. Median abnormality scores were compared using the Wilcoxon signed-rank test and performance characteristics were compared at selected cut-off thresholds between the two versions. The DICOM test library contained 2,708 participants, of whom 10.3% had a Mycobacterium tuberculosis (MTB) Positive Ultra test result. The newer software version had a significantly higher AUC than its predecessor (AUC 0.78 vs 0.76, p=0.029), and performed significantly better among people with a past history of TB (AUC 0.73 vs 0.67, p=0.003), older individuals (0.75 vs 0.77, p=0.040) and males (0.73 vs 0.76, p=0.008). The median abnormality score was significantly higher for the newer software version (0.61 vs 0.35, p<0.001) and if the newer software’s cut-off threshold was not optimized, its performance was significantly less accurate than its predecessors as the same cut-off threshold. Although INSIGHT CXR v3.9.0.1 has significantly improved performance characteristics compared to its predecessor, further studies should prospectively assess how these performance differences translate into real-world improvements during TB screening. As new CAD software versions are rolled out, the selection of a cut-off threshold must be assessed and re-calibrated to ensure the continued accuracy of CXR interpretation.