Abstract Rationale Mucus plugs (MPs) are linked to increased risk of exacerbation and mortality in obstructive lung disease and may represent a treatment target. Automated, artificial intelligence-based detection of MPs could help overcome the limitations of subjective visual assessment on computed tomography (CT). Methods Baseline CT scans from 280 participants in the COPDGene Study, selected to represent a range of GOLD stage severity, were evaluated for MPs by two radiologists (one senior, one fellow) and by an AI-based tool (LungQ, Thirona). The radiologists first evaluated the CTs independently, without the AI results, and then reviewed the scans with the AI-identified MPs displayed. Presence or absence of MPs and the total number detected per participant were compared between readers using Cohen’s kappa and the intraclass correlation coefficient (ICC). MP scores were also correlated with pulmonary function and survival. Results MPs were identified in approximately one third of CTs. Across all three readers, there was agreement on the presence or absence of MPs in 217 of 280 participants (77.5%). Cohen’s kappa values for agreement between the two radiologists, the senior radiologist and the AI tool, and the fellow and the AI tool were 0.67, 0.69, and 0.62, respectively. When comparing the total number of MPs detected, the corresponding ICCs were 0.73, 0.76, and 0.53. Among 85 participants with MPs reported by the AI, 12 (14%) and 18 (21%) were not marked by the senior radiologist and fellow, respectively, representing potential AI false positives. In participants where the AI did not segment MPs, the senior radiologist and fellow identified MPs in 26 and 28 cases, respectively, representing AI false negatives. After reviewing with AI results, the senior radiologist identified 3 participants and the fellow 6 participants with AI-detected MPs that were missed on initial visual review. Total number of MPs identified by each reader showed moderate, negative correlation with FEV1 percentage predicted (Spearman rho of -0.49, -0.34 and -0.54 for senior radiologist, fellow and AI, resepectively). Similarly, total MP counts were moderately correlated with FVC percentage predicted for each reader (-0.30, -0.20, and -0.31). The quantity of MPs labeled by readers was also significantly associated with survival (HRs 1.08, 1.03, and 1.12 for senior radiologist, fellow, and AI, respectively. Figure). Conclusions Automatic CT analysis using AI is consistent with visual assessment of MPs, is associated with poorer pulmonary function and survival, and may facilitate prompt detection, particularly for less experienced radiologists. This abstract is funded by: This work was supported by NHLBI grants U01 HL089897 and U01 HL089856 and by NIH contract 75N92023D00011.
Humphries et al. (2026) studied this question.