Radiologist and AI performance in detecting mucus plugs on chest CT.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, National Jewish Health, Denver, CO, USA. [email protected].
- Department of Radiology, National Jewish Health, Denver, CO, USA.
- Department of Radiology, School of Medicine, Duke University, Durham, NC, USA.
- Thirona, Nijmegen, The Netherlands.
- Department of Pediatric Pulmonology and Department of Radiology and Nuclear Medicine, Erasmus MC, Rotterdam, The Netherlands.
Abstract
To compare automated mucus plug detection with radiologist evaluation, assess AI-assisted detection, and examine associations of algorithm-derived airway obstruction burden with COPD severity. This retrospective analysis included 278 COPDGene participants selected across never-smokers, smokers with preserved ratio-impaired spirometry, and GOLD stages 0-4. An AI tool generated candidate plugs. Two thoracic radiologists independently identified plugs in unaided and AI-assisted sessions. A third thoracic radiologist adjudicated all candidates as the reference. Patient- and plug-level performance, observer agreement, false-positive detection, and rank-transformed associations with pulmonary function, CT parameters, and clinical severity were evaluated. Reference mucus plugs were present in 114/278 participants (41%). Patient-level AI sensitivity was lower than R1 (68.4% vs 86.8%; p < 0.001) but not different from R2 (68.4% vs 79.8%; p = 0.141), whereas specificity was similar across approaches (95.7% vs 98.8%-100.0%). Plug-level AI sensitivity was lower than both readers (34.8% vs 58.9% for R1 and 55.6% for R2; both p < 0.001). AI assistance improved readers' plug-level sensitivity (63.6-68.6%; p ≤ 0.004). AI-reader agreement was good (κ, 0.62-0.69), and inter-reader agreement improved with AI (κ, 0.70 vs 0.78; p < 0.001). These non-reference-positive AI detections (0.396/scan) were mostly marked bronchial wall thickening or partially occlusive mucus. They were independently associated with worse airflow limitation, CT emphysema, Pi10, and clinical severity after accounting for true-positive burden, improving model fit (ΔR², 0.03-0.05). AI was less sensitive than radiologists but had similar specificity for mucus plug detection. AI assistance improved reader sensitivity and agreement. Additional algorithm-detected airway abnormalities remained associated with airway disease severity. QuestionAutomated mucus plug detection on chest CT may enable standardized COPD assessment, but independent validation against radiologist assessments remains limited. FindingsArtificial intelligence showed lower sensitivity than radiologists but similar patient-level specificity on chest CT, and assistance improved sensitivity and inter-reader agreement. Clinical relevanceAutomated mucus plug analysis may support standardized airway obstruction burden assessment in COPD research and clinical trials while retaining radiologist oversight.