Quantitative CT definition of solid components in pulmonary ground-glass nodules: a pixel-level annotation-based development and internal validation study
In brief
Ground-glass-adjusted CT method identified solid nodule pixels with an AUC of 0.85
In 170 surgically resected lung nodules, a threshold adjusted to each nodule's ground-glass attenuation best distinguished expert-annotated solid pixels, with an area under the curve of 0.85. Lesion-adaptive methods performed better overall than fixed thresholds, but the study used one center and scanner; external validation is needed before clinical use.
- Journal
- Translational lung cancer research (Q1)
- Published
- 29 September 2026
- Study design
- Randomized controlled trial
- Evidence level
- Level 1, High (CEBM 1b)
- Authors
- Yuhao Chen, Nan Chen, Yuhuan Ye, Han Zhang, Wanting Xia, Kunhao Wu, et al.
- PMID
- 42999425
- DOI
- 10.21037/tlcr-2026-0841
Why clinicians should know about it
- Picked for Radiology, Radiation Oncology, Nuclear Medicine, Medical Physics and Imaging (top studies of the week, 15 November 2026): Quantitative CT definition of solid GGN components
- Picked for Pulmonary and Respiratory Medicine (top studies of the week, 15 November 2026): CT characterization of pulmonary nodules informs lung cancer diagnosis
Abstract
BACKGROUND: Solid-component assessment in pulmonary ground-glass nodules (GGNs) remains subjective, and no universally accepted computed tomography (CT) attenuation threshold exists. Fixed cutoffs may not account for inter-lesion variation in baseline ground-glass opacity (GGO) attenuation. We aimed to develop and internally validate lesion-adaptive quantitative CT definitions of the solid component using three-dimensional pixel-level expert annotations as the imaging reference standard. METHODS: This retrospective single-center methodological study included 170 unique patients with 170 surgically resected stage I lung adenocarcinoma-spectrum GGNs from March 2015 through July 2018 and each patient contributed one target lesion. All examinations were acquired on the same 256-row CT scanner, and venous-phase 1-mm images reconstructed with medium-strength deep learning image reconstruction were used for analysis. Each lesion was independently annotated by at least two experienced thoracic surgeons, with Dice-based adjudication and senior review. Patients were randomly allocated to training (n=120) and internal validation (n=50) cohorts. Nine prespecified study methods and three literature-based fixed thresholds were evaluated at pixel and lesion levels. RESULTS: The training and validation cohorts had median lesion sizes of 1.2 and 1.3 cm and median annotated solid-component proportions of 0.324 and 0.319, respectively. The GGO mean plus ratio method achieved the highest area under the receiver operating characteristic (ROC) curve [AUC, 0.848; 95% confidence interval (CI): 0.841-0.856], with sensitivity of 0.732, specificity of 0.815, and F1/Dice of 0.760. The GGO mean plus difference method yielded the highest lesion-level coefficient of determination (R2, 0.744; 95% CI: 0.499-0.889), whereas the ratio method yielded the lowest root mean square error (RMSE, 0.105) and mean absolute error (MAE, 0.088). Pairwise DeLong testing showed no statistically significant AUC difference among the three top-performing GGO-referenced methods. CONCLUSIONS: In this single-center internal validation study, lesion-adaptive GGO-referenced methods showed more favorable overall pixel-level discrimination and lesion-level agreement than universal fixed thresholds or lesion-summary estimators. These rules should be considered developmental quantitative methods and require multicenter external validation across CT platforms and reconstruction protocols before routine clinical use.
Abstract as published, via PubMed.
For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.