Repeatability of AI-quantified incidental findings on lung cancer screening CT scans in the NELSON trial

OBJECTIVES: Computed tomography (CT) scans for lung cancer screening provide the opportunity of quantifying incidental findings. We evaluated the repeatability of AI-based measurements of incidental findings using short-term repeat CT scan pairs. MATERIALS AND METHODS: AI-Rad Companion Chest CT software was applied to low-dose non-contrast CT scans from the NELSON lung cancer screening trial to measure aorta diameters, coronary artery calcium volume (CACV), vertebral height and radiodensity, and emphysema (low attenuation area percentage, LAA%). Categories (absent/present) of aortic dilatation and osteopenia, and severity (none/mild/moderate/severe) of CACV and emphysema were calculated. We included subjects who had a short-term repeat CT scan pair with a maximum interval of 120 days. We analyzed repeatability with absolute and relative differences, and agreement with the intraclass correlation coefficient (ICC) and Cohen's kappa. RESULTS: for CACV, 0.3 mm to 0.4 mm for vertebral height, and 7.6 HU for vertebral radiodensity. Median absolute difference was 0.8% for LAA%. Aorta diameters showed good (0.75 < ICC < 0.9) to excellent (ICC > 0.9) agreement. Agreement for the rest was excellent. Categorization Cohen's kappa between the first and second measurements was 0.68 for aortic dilatation, 0.63 for CACV, 0.84 for osteopenia, and 0.70 for emphysema. CONCLUSION: In a lung cancer screening cohort with short-term repeat CT, the repeatability and agreement of automated AI measurements of aortic diameters, coronary calcium, vertebral height and radiodensity, and emphysema was good to excellent. KEY POINTS: Question Can AI measurements of incidental findings beyond lung nodules be repeatably performed on pairs of lung cancer screening chest CT scans of the same subject? Findings The repeatability and agreement of AI-based measurements of aorta diameter, coronary calcium, vertebrae, and emphysema were generally excellent. Clinical relevance Measurement by radiologists of incidental findings on lung cancer screening chest CT scans imposes a high workload. AI algorithms allow for automatic measurement of incidental findings on lung cancer screening chest CT with high repeatability and agreement.