AUTHOR=Zhu Wenchao , Xu Fangyi , Lou Kaihua , Qiu Xia , Huang Dingping , Huang Shaoyu , Xie Dong , Hu Hongjie 

TITLE=The impact of Inter-observation variation on radiomic features of pulmonary nodules

JOURNAL=Frontiers in Oncology

VOLUME=Volume 15 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2025.1567028

DOI=10.3389/fonc.2025.1567028

ISSN=2234-943X

ABSTRACT=ObjectiveIn this study, we aimed to comprehensively and systematically analyze the radiomic features of pulmonary nodules and explore the influence of inter-observation variation (IOV) in segmentation regions of interest (ROI) on radiomic features, providing reference information for pulmonary nodule radiomics research.MethodSix clinicians with varying experience and expertise manually outlined ROIs for 232 pulmonary nodules, while an artificial intelligence (AI) algorithm was trained for automated segmentation. The segmentation by the most experienced cardiothoracic diagnostician (Doctor A) served as the reference standard. Inter-observer variability was assessed through diameter measurements, segmentation ROI consistency analysis, and radiomic features stability analysis.ResultsOf all radiomics features analyzed, 1071 (85.96%) demonstrated good stability (overall concordance correlation coefficient [OCCC] ≥ 0.75), with 766 (61.48%) exhibiting very good stability (OCCC ≥ 0.90). Among the eight radiomic feature types, Original _first-order, Original_GLCM, Original_GLRLM, Original_GLSZM, LOG, and wavelet features all achieved stability rates exceeding 80.00%, with 91.59% of the LOG features having good stability. The Original features showed good stability (median OCCC: 0.92-0.95, IQR: 0.12-0.19), both in the overall distribution and in the different feature categories. The median OCCC value for the LOG features (median: 0.94, IQR: 0.08) was significantly higher than that for the Wavelet features (median: 0.91, IQR: 0.13). There was no statistically significant difference in stability between the Original and LOG feature subgroups (P > 0.05). In contrast, statistically significant differences were observed between the wavelet feature subgroups (P < 0.05), with Wavelet_LLL and Wavelet_LLH transformation yielding higher stability.ConclusionSegmentation results indicated that while IOV influenced radiomic features of pulmonary nodules, most (85.96%) of the features were well stabilized and relatively unaffected. Enhancing segmentation ROI consistency helps minimize the impact of IOV on the radiomic features of pulmonary nodule images. Original and LOG features demonstrated high stability, whereas Wavelet features were more susceptible to IOV.