Open AccessRadiomics study on pulmonary infarction mimicking community‐acquired pneumonia
Author(s) -
Gao Li,
Li Yuze,
Zhai Zhenguo,
Liang Tian,
Zhang Qiang,
Xie Sheng,
Chen Huijun
Publication year - 2021
Publication title -
the clinical respiratory journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.789
H-Index - 33
eISSN - 1752-699X
pISSN - 1752-6981
DOI - 10.1111/crj.13341
Subject(s) - medicine , concordance , radiology , radiomics , pneumonia , receiver operating characteristic , community acquired pneumonia , medical diagnosis , computed tomography , pneumonia severity index , retrospective cohort study
and Objectives Pulmonary infarction (PI) shares similar symptoms and imaging presentations with community‐acquired pneumonia (CAP), which might delay diagnosis and lead to devastating consequences. Noncontrast computed tomography (CT) is the first‐line examination for the patients with the respiratory symptoms. This study aimed to investigate a radiomics method to differentiate PI from CAP using noncontrast‐enhanced CT. Methods Noncontrast‐enhanced CT images of 54 patients with PI and 64 patients with CAP were retrospectively selected. All patients were confirmed using computed tomography pulmonary angiography (CTPA). A radiomics model was built with 18 texture features that showed significant differences between PI and CAP patients. For comparison, a clinical model using clinical biomarkers and an integrated model combining the radiomics and clinical biomarkers were also generated. An experienced radiologist performed diagnoses using the noncontrast‐enhanced CT images. The parameters of the models were generated using a training dataset of 61 patients, whereas the performance of the models was evaluated using receiver operating characteristic (ROC) analysis and Harrell's concordance index (C‐index) applied to a separate validation dataset of 57 patients. Results The integrated model achieved the best performance (C‐index 0.760, sensitivity 0.703, specificity 0.867, positive predictive value [PPV] 0.826, and negative predictive value [NPV] 0.765). The radiomics model was better than both the clinical model and the radiologist's interpretations (C‐index 0.721, 0.707, 0.665, respectively; sensitivity 0.667, 0.630, 0.593; specificity 0.800, 0.785, 0.733; PPV 0.750, 0.739, 0.667; and NPV 0.727, 0.706, 0.667). Conclusions Radiomics features generated from noncontrast‐enhanced CT images allow PI to be differentiated from CAP with considerable accuracy. The radiomics‐based method could provide useful information in clinical practice.