Performance of a deep-learning algorithm for referable thoracic abnormalities on chest radiographs: A multicenter study of a health screening cohort

• DOI: 10.1371/journal.pone.0246472
• Publication Date: 2021-02-19T19:57:16Z
• Authors: Eun Young Kim; Young Jae Kim; Won-Jun Choi; Gi Pyo Lee; Ye Ra Choi; Kwang Nam Jin; Young Jun Cho
• Keywords: Adult; Male; Science; Q; R; Middle Aged; 3. Good health; Deep Learning; ROC Curve; Medicine; Humans; Multicenter Studies as Topic; Female; Radiography, Thoracic; Algorithms; Research Article; Aged; Retrospective Studies

Purpose This study evaluated the performance of a commercially available deep-learning algorithm (DLA) (Insight CXR, Lunit, Seoul, South Korea) for referable thoracic abnormalities on chest X-ray (CXR) using consecutively collected multicenter health screening cohort. Methods and materials A consecutive cohort participants who underwent both CXR computed tomography (CT) within 1 month was retrospectively from three institutions’ care clinics (n = 5,887). Referable were defined as any radiologic findings requiring further diagnostic evaluation or management, including DLA-target lesions nodule/mass, consolidation, pneumothorax. We DLA area under receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity ground truth based CT (CT-GT). In addition, CT-GT-positive cases, independent radiologist readings performed clear visible (when more than two radiologists called) (at least one CXR-GTs (clear CXR-GT CXR-GT, respectively) to evaluate DLA. Results Among 5,887 subjects (4,329 males; mean age 54±11 years), found in 618 (10.5%) CT-GT. observed 223 (4.0%), nodule/mass 202 (3.4%), consolidation 31 (0.5%), pneumothorax (<0.1%), DLA-non-target 409 (6.9%). For CT-GT, showed an AUC 0.771 (95% confidence interval [CI], 0.751–0.791), sensitivity 69.6%, 74.0%. Based prevalence decreased, with 405 (6.9%) 227 (3.9%) respectively. The increased significantly when CXR-GTs, 0.839 CI, 0.829–0.848), 82.7%, s 73.2% 0.872 0.863–0.880, P <0.001 comparison GT-CT vs. CXR-GT), 83.3%, 78.8% CXR-GT. Conclusion provided fair-to-good stand-alone detection varied according different methods truth.