Concordance rate of radiologists and a commercialized deep-learning solution for chest X-ray: Real-world experience with a multicenter health screening cohort

• Published in: PloS one Vol. 17; no. 2; p. e0264383
• Main Authors: Kim, Eun Young, Kim, Young Jae, Choi, Won-Jun, Jeon, Ji Soo, Kim, Moon Young, Oh, Dong Hyun, Jin, Kwang Nam, Cho, Young Jun
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 24.02.2022; Public Library of Science (PLoS)
• Subjects: Abnormalities; Aged; Algorithms; Archives & records; Artificial intelligence; Biomedical engineering; Chest; Clinical medicine; Cohort Studies; Commercialization; Decision making; Decision support systems; Deep Learning; Discordance; Diseases; Female; Health care; Hospitals; Humans; Integration; Learning; Lesions; Lung cancer; Machine learning; Male; Mass Screening; Medical diagnosis; Medical schools; Medical screening; Medicine; Middle Aged; Physicians; Pneumothorax; Practice; Predictive Value of Tests; Radiography; Radiography, Thoracic - methods; Radiography, Thoracic - standards; Radiologists; Radiology; Reading; Retrospective Studies; Statistics; Technology application; Thorax; South Korea
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0264383

Lunit INSIGHT CXR (Lunit) is a commercially available deep-learning algorithm-based decision support system for chest radiography (CXR). This retrospective study aimed to evaluate the concordance rate of radiologists and Lunit for thoracic abnormalities in a multicenter health screening cohort. We retrospectively evaluated the radiology reports and Lunit results for CXR at several health screening centers in August 2020. Lunit was adopted as a clinical decision support system (CDSS) in routine clinical practice. Subsequently, radiologists completed their reports after reviewing the Lunit results. The DLA result was provided as a color map with an abnormality score (%) for thoracic lesions when the score was greater than the predefined cutoff value of 15%. Concordance was achieved when (a) the radiology reports were consistent with the DLA results ("accept"), (b) the radiology reports were partially consistent with the DLA results ("edit") or had additional lesions compared with the DLA results ("add"). There was discordance when the DLA results were rejected in the radiology report. In addition, we compared the reading times before and after Lunit was introduced. Finally, we evaluated systemic usability scale questionnaire for radiologists and physicians who had experienced Lunit. Among 3,113 participants (1,157 men; mean age, 49 years), thoracic abnormalities were found in 343 (11.0%) based on the CXR radiology reports and 621 (20.1%) based on the Lunit results. The concordance rate was 86.8% (accept: 85.3%, edit: 0.9%, and add: 0.6%), and the discordance rate was 13.2%. Except for 479 cases (7.5%) for whom reading time data were unavailable (n = 5) or unreliable (n = 474), the median reading time increased after the clinical integration of Lunit (median, 19s vs. 14s, P < 0.001). The real-world multicenter health screening cohort showed a high concordance of the chest X-ray report and the Lunit result under the clinical integration of the deep-learning solution. The reading time slight increased with the Lunit assistance.