Application of deep learning algorithm to detect COVID-19 pneumonia in chest X-ray

Jang, Se Bum; Chung, Han Sol; Park, Sin-Yul

J Korean Soc Emerg Med > Volume 32(3); 2021 > Article

Original Article

Journal of The Korean Society of Emergency Medicine 2021;32(3): 249-256.

흉부 X선에서 COVID-19 폐렴 검출을 위한 DL 알고리즘의 적용

장세범

, 정한솔

, 박신률

영남대학교 의과대학 응급의학교실

Application of deep learning algorithm to detect COVID-19 pneumonia in chest X-ray

Se Bum Jang

, Han Sol Chung

, Sin-Yul Park

Department of Emergency Medicine, Yeungnam University College of Medicine, Daegu, Korea

Correspondence

Han Sol Chung ,Tel: 053-620-3193, Fax: 053-623-8030, Email: ssolkun@gmail.com,

Received: October 20, 2020; Revised: November 6, 2020 Accepted: November 6, 2020. Published online: June 30, 2021.

ABSTRACT

Objective:
This study evaluated the deep learning (DL) algorithm performance to detect lesions that suggest pneumonia in chest X-rays (CXR) of suspected coronavirus disease 2019 (COVID-19) patients.
Method:
This retrospective study included consecutive patients who visited a screening clinic in Daegu, and were suspected to be afflicted with the COVID-19 during the COVID-19 epidemic. CXR were analyzed using the commercial artificial intelligence product that provides free online DL algorithms to the public for COVID-19. Computerized tomography was used as the standard reference. Performance of the DL algorithm was evaluated by the sensitivity and specificity, and results were compared to the CXR records of emergency physicians (EP) in charge of the actual screening triage clinic during the COVID-19 epidemic.
Results:
Totally, 114 patients were evaluated, of which 38 patients were positive for COVID-19. In 85 CXRs examined (36 COVID-19 and 49 non-COVID-19) with findings of pneumonia in computerized tomography, the DL algorithm showed significantly higher sensitivity as compared to the EP (DL, 98.8% [93.6%-99.9%] vs. EP, 85.9% [76.6%-92.5%]; P<0.01). Moreover, the DL algorithm showed significantly higher sensitivity for detecting CXRs with COVID-19 pneumonia, as compared to the EP (DL, 100.0% [90.3%-100%] vs. EP, 91.7% [77.5%-98.3%]; P=0.08).
Conclusion:
We conclude that for examining the CXR of patients with suspected COVID-19, sensitivity of the DL algorithm is superior than the EP for detecting lesions suggesting pneumonia. Thus, the application of the DL algorithm is potentially useful in screening triage clinics to detect COVID-19 pneumonia.

Key words: Deep learning; COVID-19; Screening triage; COVID-19 diagnostic testing; Radiography