Ann Transl Med . Deep learning for differentiating novel coronavirus pneumonia and influenza pneumonia

[tetano]

Ann Transl Med


. 2021 Jan;9(2):111.
doi: 10.21037/atm-20-5328.
Deep learning for differentiating novel coronavirus pneumonia and influenza pneumonia


Min Zhou ^{1 2} , Dexiang Yang ³ , Yong Chen ⁴ , Yanping Xu ^{1 2} , Jin-Fu Xu ⁵ , Zhijun Jie ⁶ , Weiwu Yao ⁷ , Xiaoyan Jin ⁸ , Zilai Pan ⁹ , Jingwen Tan ⁴ , Lan Wang ⁴ , Yihan Xia ⁴ , Longkuan Zou ¹⁰ , Xin Xu ¹⁰ , Jingqi Wei ¹⁰ , Mingxin Guan ¹⁰ , Fuhua Yan ⁴ , Jianxing Feng ¹⁰ , Huan Zhang ⁴ , Jieming Qu ^{1 2}



Affiliations

• PMID: 33569413
• PMCID: PMC7867927
• DOI: 10.21037/atm-20-5328

Abstract

Background: Chest computed tomography (CT) has been found to have high sensitivity in diagnosing novel coronavirus pneumonia (NCP) at the early stage, giving it an advantage over nucleic acid detection during the current pandemic. In this study, we aimed to develop and validate an integrated deep learning framework on chest CT images for the automatic detection of NCP, focusing particularly on differentiating NCP from influenza pneumonia (IP).
Methods: A total of 148 confirmed NCP patients [80 male; median age, 51.5 years; interquartile range (IQR), 42.5-63.0 years] treated in 4 NCP designated hospitals between January 11, 2020 and February 23, 2020 were retrospectively enrolled as a training cohort, along with 194 confirmed IP patients (112 males; median age, 65.0 years; IQR, 55.0-78.0 years) treated in 5 hospitals from May 2015 to February 2020. An external validation set comprising 57 NCP patients and 50 IP patients from 8 hospitals was also enrolled. Two deep learning schemes (the Trinary scheme and the Plain scheme) were developed and compared using receiver operating characteristic (ROC) curves.
Results: Of the NCP lesions, 96.6% were >1 cm and 76.8% were of a density <-500 Hu, indicating them to have less consolidation than IP lesions, which had nodules ranging from 5-10 mm. The Trinary scheme accurately distinguished NCP from IP lesions, with an area under the curve (AUC) of 0.93. For patient-level classification in the external validation set, the Trinary scheme outperformed the Plain scheme (AUC: 0.87 vs. 0.71) and achieved human specialist-level performance.
Conclusions: Our study has potentially provided an accurate tool on chest CT for early diagnosis of NCP with high transferability and showed high efficiency in differentiating between NCP and IP; these findings could help to reduce misdiagnosis and contain the pandemic transmission.

Keywords: Deep learning; chest computed tomography (CT); influenza pneumonia (IP); novel coronavirus pneumonia (NCP).