Objectives: To evaluate the diagnostic accuracy of radiomics method and frozen sections (FS) for the pathological classification of peripheral lung adenocarcinoma manifesting as ground-glass nodules (GGNs) in computer tomography (CT).
Materials and methods: A dataset of 831 peripheral lung adenocarcinoma manifesting as GGNs in CT were divided into two cohorts: training cohort (n = 581) and validation cohort (n = 250). Combined with clinical features, the radiomics classifier was trained and validated to distinguish the pathological classification of these nodules. FS diagnoses in the validation cohort were collected. Diagnostic performance of the FS and radiomics methods was compared in the validation cohort. The predictive factors for the misdiagnosis of FS were determined via univariate and multivariate analyses.
Results: The accuracy of radiomics method in the training and validation cohorts was 72.5 % and 68.8 % respectively. The overall accuracy of FS in the validation cohort was 70.0 %. The concordance rate between FS and final pathology when FS had a different diagnosis from radiomics classifier was significantly lower than when FS had the same diagnosis as radiomics classifier (46 vs. 87 %, P < 0.001). Univariate and Multivariate analyses showed that different diagnosis between FS and radiomics classifier was the independent predictive factor for the misdiagnosis of FS (OR: 7.46; 95%CI: 4.00-13.91; P < 0.001).
Conclusions: Radiomics classifier predictions may be a reliable reference for the classification of peripheral lung adenocarcinoma manifesting as GGNs when FS cannot provide a timely diagnosis. Intraoperative diagnoses of the cases where FS had a different diagnosis from radiomics method should be considered cautiously due to the higher misdiagnosis rate.
Keywords: Adenocarcinoma of lung; Diagnostic errors; Frozen sections; Risk factors; Spiral computed; Thoracic surgery; Tomography.
Copyright © 2019 Elsevier B.V. All rights reserved.