Deep learning-based automated segmentation and volumetric measurement improve inter-observer consistency and reduce false-positive findings in Lung-RADS assessment of solid pulmonary nodules

To systematically evaluate inter-observer variability in nodule measurement and Lung CT Screening Reporting and Data System (Lung-RADS) classification for solid pulmonary nodules, and to determine whether semiautomated diameter and volumetric measurements improve reproducibility and clinical decision consistency. Three radiologists retrospectively evaluated solid pulmonary nodules detected during lung cancer screening between January 2021 and December 2023 that were categorized as Lung-RADS 2, 3, 4A, or 4B. Volumetric data and semiautomated mean diameters were generated using a deep learning-based automatic pulmonary nodule segmentation algorithm built on the nnU-Net v2 framework. Major inconsistency was defined as a discrepancy of two or more Lung-RADS categories. Inter-observer agreement for nodule measurements was assessed using intraclass correlation coefficients (ICC) and Bland-Altman analysis, while agreement in Lung-RADS categorization was evaluated using Cohen κ for paired observers and Fleiss κ among three observers. For all observer pairs, ICC for all three measurement methods were greater than 0.90. Semiautomated measurements showed smaller bias and narrower 95% limits of agreement compared with manual measurements. Cohen κ values were 0.522–0.759 for manual measurements, 0.825–0.874 for semiautomated diameter measurements, and 0.882–0.961 for semiautomated volumetry. Fleiss κ values were 0.501 (0.422–0.567), 0.773 (0.701–0.845), and 0.868 (0.794–0.942), respectively. Volumetric measurements reduced major management discrepancies by approximately two-thirds. Using semiautomated volumetric assessment, all three observers classified a higher proportion of nodules as Lung-RADS category 2 (52.2%, 47.8%, and 44.4%, respectively). Volumetric assessment prolonged the mean follow-up interval, increasing from 6.9 months with manual measurements to 7.8 months. Compared with manual measurements, semiautomated assessments improved the consistency of nodule measurements. CT-based volumetric assessment classified a greater proportion of nodules into lower Lung-RADS categories. Semi-automated volumetric measurement reduces the false-positive rate of nodule detection. Not applicable.